Towards Zero and Few-shot Knowledge-seeking Turn Detection in Task-orientated Dialogue Systems

TL;DR

This paper introduces REDE, a novel method for zero and few-shot turn detection in task-oriented dialogue systems, capable of identifying out-of-scope user requests with minimal data and quick adaptation.

Contribution

The paper presents REDE, a new adaptive representation learning approach that enables effective zero-shot and few-shot turn detection in dialogue systems, reducing reliance on large annotated datasets.

Findings

REDE achieves competitive performance on DSTC9 data.

REDE adapts quickly with fewer than 3K parameter updates.

REDE outperforms existing methods in low-data scenarios.

Abstract

Most prior work on task-oriented dialogue systems is restricted to supporting domain APIs. However, users may have requests that are out of the scope of these APIs. This work focuses on identifying such user requests. Existing methods for this task mainly rely on fine-tuning pre-trained models on large annotated data. We propose a novel method, REDE, based on adaptive representation learning and density estimation. REDE can be applied to zero-shot cases, and quickly learns a high-performing detector with only a few shots by updating less than 3K parameters. We demonstrate REDE's competitive performance on DSTC9 data and our newly collected test set.