Core Concepts
BootTOD proposes a self-bootstrapping framework for task-oriented dialogue representations, outperforming contrastive methods.
Abstract
Abstract: BootTOD introduces a self-bootstrapping framework for task-oriented dialogue representations, addressing limitations of contrastive methods.
Introduction: Previous unsupervised pre-training models for task-oriented dialogues have relied on contrastive learning, facing challenges in selecting true positives and hard negatives.
Model: BootTOD aligns context and context+response representations, dismissing the need for contrastive pairs, and models the one-to-many diversity of human conversations.
Experiment: BootTOD achieves consistent improvements over strong TOD baselines in various dialogue tasks, showcasing its generalization capability.
Qualitative Analysis: Ablation study shows the importance of alignment layers and max response length in enhancing performance.
Non-Contrastive Methods Comparison: BootTOD outperforms FutureTOD in intent recognition, dialogue state tracking, dialogue act prediction, and response selection tasks.
Stats
Pre-trained language models successful in many scenarios.
BootTOD outperforms strong TOD baselines on diverse downstream dialogue tasks.
Quotes
"BootTOD aligns context and context+response representations and dismisses the requirements of contrastive pairs."
"Experimental results show that BootTOD outperforms strong TOD baselines on diverse downstream dialogue tasks."