SUMBT+LaRL: Effective Multi-Domain End-to-End Neural Task-Oriented Dialog System

The recent advent of neural approaches for developing each dialog component in task-oriented systems has remarkably improved, yet optimizing the overall system performance remains a challenge. Besides, previous research on modeling complicated multi-domain goal-oriented dialogs end-to-end fashion been limited. In this paper, we present an effective trainable SUMBT+LaRL that incorporates two strong models and facilitates them to be fully differentiable. Specifically, SUMBT+ estimates user-acts as well belief states, LaRL latent action spaces generates responses given estimated contexts. We emphasize training framework three steps significantly stably increase success rates: separately pretraining LaRL, fine-tuning entire system, then reinforcement learning policy. also introduce new reward criteria policy training. Then, discuss experimental results depending different evaluation methods. Consequently, our model achieved state-of-the-art rate 85.4% corpus-based evaluation, comparable 81.40% simulator-based provided by DSTC8 To best knowledge, work is first comprehensive study modularized E2E from task success.