Multi-Domain Dialog State Tracking based on Machine Reading Comprehension

With the development of big data and deep learning technology, the goal of creating an automatic human-machine dialog system is no longer an illusion. However, obtaining correct dialog label is very challenging due to expensive purchase cost and time cost. Through the transfer of knowledge, we proposed RC_DST which enables the dialog state tracking model to infer the state of dialog between different domains accurately and improves the accuracy in a new domain without or with a few of labeled data. In this paper, we utilizes XLNet which performs well in processing long texts to encode the dependency between the dialog context and slot semantics. Meanwhile, we expands machine reading comprehension to non-categorical and categorical slots in different ways. Extensive experiment results show that our method achieves competitive results in a new domain with zero samples compared with exclusive training data in this domain.