DS-TOD: Efficient Domain Specialization for Task Oriented Dialog

TL;DR

This paper introduces DS-TOD, a domain-specific pretraining framework for task-oriented dialog that uses domain extraction, resources, and adapters to improve performance across multiple domains with efficiency.

Contribution

The work proposes a novel domain specialization method for pretrained language models in TOD, utilizing domain-specific resources and lightweight adapters for improved multi-domain performance.

Findings

DS-TOD improves dialog state tracking and response retrieval across five domains.

Lightweight adapters match full fine-tuning performance in single domains.

Adapters offer computational efficiency and better multi-domain results.

Abstract

Recent work has shown that self-supervised dialog-specific pretraining on large conversational datasets yields substantial gains over traditional language modeling (LM) pretraining in downstream task-oriented dialog (TOD). These approaches, however, exploit general dialogic corpora (e.g., Reddit) and thus presumably fail to reliably embed domain-specific knowledge useful for concrete downstream TOD domains. In this work, we investigate the effects of domain specialization of pretrained language models (PLMs) for TOD. Within our DS-TOD framework, we first automatically extract salient domain-specific terms, and then use them to construct DomainCC and DomainReddit -- resources that we leverage for domain-specific pretraining, based on (i) masked language modeling (MLM) and (ii) response selection (RS) objectives, respectively. We further propose a resource-efficient and modular domain specialization by means of domain adapters -- additional parameter-light layers in which we encode the domain knowledge. Our experiments with prominent TOD tasks -- dialog state tracking (DST) and response retrieval (RR) -- encompassing five domains from the MultiWOZ benchmark demonstrate the effectiveness of DS-TOD. Moreover, we show that the light-weight adapter-based specialization (1) performs comparably to full fine-tuning in single domain setups and (2) is particularly suitable for multi-domain specialization, where besides advantageous computational footprint, it can offer better TOD performance.