Research On New Domain Adaptation Strategy Of Task-oriented Dialog System

In application to new fields,task-oriented dialogue systems based on deep learning often require a large number of labeled dialogues covering almost all possible dialogue processes.However,correct dialogue labeling is very challenging due to expensive purchase costs,long time periods,and regulatory issues.Therefore,when the dialogue task is changed,how to quickly and accurately extend the corresponding model to a new field with less labeled data and reduce the dependence on the labeled data in the field is very important and worth exploring in practical applications.This paper focuses on two key technologies for bridging the communication barriers between humans and computers in dialogue systems,semantic parsing and dialogue state tracking,and studies new-field adaptation strategies in scenarios with limited labeled data.We propose solutions from two perspectives:using unlabeled data in the new field and transfer knowledge to the new field.Aiming at the characteristics of semantic parsing:diversity,professional and poor transferability of meaning representation,this paper utilizes the relatively easy-to-obtain and unlabeled natural language utterances in the new field under the sequence-to-sequence framework,so that the model can still show good performance in the case of less labeled data in the new field of semantic parsing.We propose a semi-supervised semantic parsing method by exploiting unlabeled natural utterances in a novel multitask learning framework.Two strategies are proposed.The first one takes entity sequences as training targets to improve the representations of encoder and reduce entity-mistakes in prediction.The second one extends Mean Teacher to sequence-to-sequence model and generates more target-side data to improve the generalizability of decoder network.Experiments demonstrate that our proposed methods significantly outperform the supervised baseline and achieve more impressive improvement than previous methods.In view of the strong readability and transferability of the dialog state,this paper works on improving the model transferability,so that the dialog state tracking model can accurately estimate the dialog state among different fields,and expanding to new fields with no labeled data or very few labeled data.Few new areas of dialogue.This paper utilizes the advantages of XLNet to process long texts,encodes the dependency between the dialogue context and slot semantics,and expands machine reading comprehension to non-categorical and categorical slots in different ways.Experiments show that the model can obtain competitive results in zero-shot settings with training data just in this field.