UBAR: Towards Fully End-to-End Task-Oriented Dialog System with GPT-2

This paper presents our task-oriented dialog system UBAR which models dialogs on a session level. Specifically, is acquired by fine-tuning the large pre-trained unidirectional language model GPT-2 sequence of entire composed user utterance, belief state, database result, act, and response every turn. Additionally, evaluated in more realistic setting, where its context has access to utterances all content it generated such as states, acts, responses. Experimental results MultiWOZ datasets show that achieves state-of-the-art performances multiple settings, improving combined score generation, policy optimization, end-to-end modeling 4.7, 3.5, 9.4 points respectively. Thorough analyses demonstrate session-level training formulation are essential for operate fully real life. We also examine transfer ability new domains with limited data provide visualization case study illustrate advantages