Reproduce “Zero-Shot Dialogue State Tracking via Cross-Task Transfer” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhaojiang Lin, Bing Liu, Andrea Madotto, Seungwhan Moon, Paul Crook, Zhenpeng Zhou, Zhiguang Wang, Zhou Yu and 3 more

Researchers from HKUST and Facebook developed TransferQA, a framework that reformulates Dialogue State Tracking (DST) as a generative Question Answering (QA) task using a unified text-to-text model. This method enables zero-shot DST by leveraging large QA datasets and effectively handling "none" values, achieving higher accuracy than previous approaches on MultiWoz and SGD benchmarks.

Key insights

  • Transferring knowledge from diverse, large-scale general Question Answering (QA) datasets offers superior generalization capabilities for zero-shot Dialogue State Tracking (DST) compared to traditional cross-domain DST transfer.
  • A unified generative model architecture effectively handles both extractive and multi-choice QA formats, which promotes better knowledge sharing and simplifies the process for managing both non-categorical and categorical DST slots.
  • Explicitly training models to identify "unanswerable questions" (using NQS and CT) is crucial for accurate "none" slot prediction, which significantly enhances the robustness and practical utility of zero-shot DST systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.