A Unified Generation Approach for Robust Dialogue State Tracking.

While dialogue state tracking by generation-based approaches allows for better scalability and generalization, they suffer from two major limitations. First, most generation-based models adopt a multi-task learning framework that may cause gradient conflicts and low training efficiency. Second, since the dialogue state of the previous turn is usually taken as an input for the current turn, there exists inconsistency between training and inference, which is identified as turn-level exposure bias. To address the first limitation, we propose the idea of state-transition sequence and transform multi-task learning into a single generation task. To alleviate turn-level exposure bias, we propose a slot-perturb strategy to reduce the over-reliance on the previous dialogue state. Experimental results show that our method achieves a new state of the art on the MultiWOZ 2.4 dataset and performs competitively on MultiWOZ 2.1. Besides, we demonstrate that the unified generation framework with slot-perturb improves the convergence speed and relieves error accumulation.