Research On Multi-Domain Dialogue State Tracking

Users interact with the task-oriented dialogue systems by multiple rounds of dialogues,and with the increase of the number of interaction rounds,the system can gradually clarify and fulfill the requirements put forward by the user.As an integral part of a task-based dialogue system,accurate and efficient dialogue state tracking performance is essential for updating the internal state of the dialogue system and formulating dialogue strategies.As a language understanding task based on contextual,the goal of the dialogue state tracking task is to extract the user’s goals and intentions from the content of each dialogue round,and represent them in the form of a compact set of dialogue states,i.e.,a set of(domain-slot)pairs.The dialogue states can provide an important reference for the system to select system actions and response discourse.With the continuous improvement of peoples’ s living standards,people prefer to accomplish different domain needs in one system through multiple rounds of interaction.For example,users can book hotels,restaurants and cabs in order,which is the task of multidomain dialogue state tracking.The research of multi domain task-based dialogue system and its related technologies has attracted extensive attention of scholars.Regarding that the performance of dialogue state tracking is closely related to the quality of system response generated by dialogue system,this thesis focuses on the task of multi domain dialogue state tracking.To alleviate the data sparsity caused by the excessive(domain-slot)pairs and enhance the implicit inference capability of the model,a multi-domain dialogue state tracking model(Integrate Relevant Information Dialogue State Tracker,IRI-DST)based on the incorporation of slot association information is proposed in this thesis.Based on the encoder-decoder architecture,a stacked correlation information module is designed to fuse the correlation information between each(domain-slot)pair,so that each slot can learn not only the information in the dialogue context but also the information between related slots.Finally,the GRU unit with a soft gating-based replication mechanism is proposed to generate the corresponding slot values from the dialogue context or vocabulary,thus improving the model’s ability to generate unknown slot values.Experimental results on public datasets demonstrate that the model has higher joint accuracy than other baseline models.In addition,to solve the slot misconnection problem caused by utilizing global knowledge information and to jointly optimize the encoding and decoding process,a multi-domain dialogue state tracking model(Dialogue State Graph Dialogue State Tracker,DSG-DST)based on incorporating dialogue state graphs is proposed in this thesis.This thesis applies the idea of heterogeneous graphs to construct dialogue state graphs specifically for all(domainslot)pairs of a single sample,treating slots as different classes of nodes regarding different domains,and using a hierarchical attention mechanism to fuse information from within and outside the domain for slots.The model also uses BERT as a slot-value decoder to share encoder parameters for joint optimization of the model.Experimental results on public datasets show that the model has better performance.Finally,a multi-domain dialogue state tracking demonstration system is designed in this thesis.The system can generate dialogue states based on dialogue data provided by users through direct input of dialogue text or uploading dialogue files,and return the generated dialogue states of each dialogue round to the front-end page for presentation.