Research On Technologies And Its Generalization In Multi-domain Task-oriented Dialogue State Tracking

The development of society and the technology makes the human-machine interaction system becoming a focus in research,and dialogue system is an important part of the human-machine interaction system.Task-oriented dialogue is one of the most widely used technologies in dialogue systems.With the promotion of application in system dialogue,the single-domain dialogue system can no longer meet actual needs.Research on multi-domain dialogue systems and relevant technologies has gradually become the focus of current research and application.And the migration and generalization of cross-language capabilities of the model in multi-domain dialogue systems have become a new research direction which have many practical application scenarios with the advancement of current internationalization.This paper analyzes and summarizes the existing public datasets for task-oriented dialogue.And this paper introduces and analyzes the two commonly used multi-domain datasets in detail.Besides,this paper gives a complete description of their generation methods.And this paper introduces the commonly used evaluation indicators for task-oriented dialogues and the most important indicators,joint accuracy,and its calculation methods.And this paper corrects some of the wrong data in the commonly used public datasets in semi-automatic method which is "manual & rule",when this paper runs a model in that data and analyzes the bad cases,so a high-quality dataset was generated.This paper researches and explores the key technology of dialogue state tracking in task-oriented dialogue systems.This paper improves and designs the encode-decode structure of existing dialogue state tracking model for multi-domain.This paper re-designs the pre-process and post-process from the aspect of data representation,information acquisition methods and pre-training task which improvs the ability in understanding of the semantics continuously expressed by the user-agent during the human-machine interaction process because this re-design merges multi-domain information into the word token of slot and extracts it separately from the results.Compared with the baseline model,the model improves evaluation indicators such as joint accuracy,and it gives a good idea for subsequent related model design.In this paper,a model transfer generalization learning task and data processing method suitable for task-based dialogues are designed in view of the tasks of cross-language transfer generalization.The model transfer generalization ability is improved from the perspectives of joint learning,corpus mapping,etc.This paper runs experiments in the cross-language generalization method of multi-domain task dialogue status tracking.This paper takes a further processing of the data from the perspective of statistics and rules according to the features of the data,and it has achieved good results in the main evaluation indicators.