Research And Application On Schema-guided Dialogue State Tracking

A task-oriented dialogue system can help users complete tasks in specific domains with natural language.Dialogue state tracking is the core component of the system.In recent years,the task-oriented dialogue system is changing from a single-domain system to a multi-domain system.Making good use of the in-domain knowledge and helping the model adapt to a new domain is worth studying.Meanwhile,the advent of word-level dialogue state tracking has avoided using an extra representation to describe the intermediate result and lowered the possibility of information loss.This paper focuses on schema-guided dialogue state tracking,which is a word-level multi-domain dialogue state tracking task.The key difference between the traditional dialogue state tracking task and the schema-guided one is the use of the schema and the natural language description of the schema elements,which makes it possible for the model to transfer knowledge to an unseen domain and achieves zero-shot dialogue state tracking.Despite their success,existing methods still suffer from two weaknesses:(1)Current methods fail to exploit the dialogue history.(2)Current methods fail to account for the differences between subtasks and treat them independently without considering their relation.To address these issues,we propose a two-stage framework for schema-guided dialogue state tracking with selected dialogue history.The main contributions are as follows:(1)We propose a dialogue history selection method for schema-guided dialogue state tracking,selecting the most related utterance to a specific subtask from the dialogue history.This method allows other modules to explore more dialogue history for longer conversations,which improves the model's performance.(2)We propose a two-stage framework for schema-guided dialogue state tracking,designing models for task-specific schema elements with Ro BERTa-based multiple-choice machine reading comprehension,Longformer-based span-based machine reading comprehension,and considering the relationship between the subtasks,which makes full use of the schema element and improves the model's performance.(3)We design a task-oriented dialogue system to complete both the data entry task and the information extraction task through human-machine dialogue.Due to the poor scalability of the current system,it is hard to meet the demand of extracting new attributes without changing the structure of the model.The proposed schema-guided dialogue state tracking method is applied,allowing the training of newly added attributes without changing the structure of the model.Current speech recognition and dialogue system technology suffer from cascading errors.A modification stage is added after the initial input stage.With the help of a BERT-based sequence labeling model,users are allowed to change the attribute with voice.Experiments are conducted on two different scenarios.The result shows that the system is effective and scalable.