Research On Dialogue State Tracking In Task-oriented Dialogue Systems

With the rapid development of artificial intelligence technology,task-oriented dialogue systems have been widely applied in the field of human-computer interaction.Dialogue state tracking(DST)is an indispensable component of task-oriented dialogue systems and a crucial link in understanding user intentions.In recent years,many DST models have been proposed,achieving satisfactory results.However,most current DST models predict the value of each slot individually,without considering the correlations among slots.Moreover,the errors in the prediction of the value of each slot at the current dialogue turn of these models are easily carried over to the next turn and are unlikely to be revised in the next turn,resulting in error propagation.To address these issues,this thesis presents the following research work and innovation points:(1)To address the issue of ignored correlations among slots,this thesis proposes an efficient slot correlation learning network for multi-domain dialogue state tracking(SCDST).The model mainly consists of four parts: dialogue context encoding,specific slot information extraction,slot self-attention,and slot value matching.First,the dialogue context encoding part utilizes the pre-trained language model BERT to encode the dialogue context,slot names,and corresponding slot values,obtaining the token vector representations.Then,the cross multi-head attention module in the specific slot information extraction part calculates and fuses the attention between the dialogue context embedding,slot names embedding,and corresponding slot values embedding,extracting relevant features and providing them to other modules to fully extract feature information for each slot based on the dialogue context.Moreover,the slot self-attention part learns the correlations among slots.Finally,the slot value matching part predicts the value of each slot.The model has good performance in scalability because it does not need any hand-crafted features or prior knowledge.Experiments are conducted on Multi WOZ 2.0,Multi WOZ 2.1,and Multi WOZ 2.4 and results indicate that SC-DST outperforms previous state-of-the-art works,achieving new state-of-the-art performances with 55.14%,57.22%,and 76.93% joint goal accuracy,respectively,which has a significant improvement(0.61%,0.86%,and 3.31%)over the previous best results.(2)To address the error propagation issue,this thesis proposes the revisable state prediction for dialogue state tracking(RSP-DST),which constructs a two-stage slot value prediction process consisting of an original dialogue state prediction and a revising dialogue state prediction.Specifically,the original prediction process jointly models the dialogue context and previous dialogue state to predict the original dialogue state of the current dialogue turn.Then,in the revising dialogue state prediction process,the model jointly models the dialogue context and the original dialogue state to avoid the errors existing in the original dialogue state carried over to the next dialogue turn,alleviating the error propagation.Experiments are conducted on Multi WOZ 2.0,Multi WOZ 2.1,and Multi WOZ 2.4 and results indicate that RSP-DST outperforms previous state-of-the-art works,achieving new state-of-the-art performances with 56.35%,58.09%,and 75.65%joint goal accuracy,respectively,which has a significant improvement(2.15%,1.73%,and 2.03%)over the previous best results.