Effective Sequence-to-Sequence Dialogue State Tracking

TL;DR

This paper systematically investigates how sequence-to-sequence models can be optimized for dialogue state tracking, highlighting the importance of pre-training objectives and context representations for improved performance.

Contribution

It introduces the use of span prediction-based pre-training, like Pegasus, for dialogue state tracking, showing its effectiveness over traditional methods.

Findings

Masked span prediction outperforms auto-regressive modeling.

Pre-training on summarization tasks benefits dialogue state tracking.

Recurrent context representations struggle with error recovery.

Abstract

Sequence-to-sequence models have been applied to a wide variety of NLP tasks, but how to properly use them for dialogue state tracking has not been systematically investigated. In this paper, we study this problem from the perspectives of pre-training objectives as well as the formats of context representations. We demonstrate that the choice of pre-training objective makes a significant difference to the state tracking quality. In particular, we find that masked span prediction is more effective than auto-regressive language modeling. We also explore using Pegasus, a span prediction-based pre-training objective for text summarization, for the state tracking model. We found that pre-training for the seemingly distant summarization task works surprisingly well for dialogue state tracking. In addition, we found that while recurrent state context representation works also reasonably well, the model may have a hard time recovering from earlier mistakes. We conducted experiments on the MultiWOZ 2.1-2.4, WOZ 2.0, and DSTC2 datasets with consistent observations.