Zero-Shot Transfer Learning with Synthesized Data for Multi-Domain Dialogue State Tracking

• DOI: 10.18653/v1/2020.acl-main.12
• Publication Date: 2020-07-29T14:14:43Z
• Authors: Giovanni Campagna; Agata Foryciarz; Mehrad Moradshahi; Monica Lam
• Keywords: Transfer of learning; Zero (linguistics); Tracking (education); Training set; Labeled data

Zero-shot transfer learning for multi-domain dialogue state tracking can allow us to handle new domains without incurring the high cost of data acquisition. This paper proposes zero-short technique where in-domain training are all synthesized from an abstract model and ontology domain. We show that augmentation through improve accuracy zero-shot both TRADE BERT-based SUMBT on MultiWOZ 2.1 dataset. with only reach about 2/3 obtained full art average across by 21%.