[Paper Review] End-to-end Task-oriented Dialogue: A Survey of Tasks, Methods, and Future Directions

Libo Qin, Wenbo Pan+6 more|arXiv (Cornell University)|Nov 15, 2023

Topic Modeling4 citations

TL;DR

This paper presents the first comprehensive survey of end-to-end task-oriented dialogue (EToD), introducing a unified taxonomy of modularly and fully end-to-end approaches, analyzing key trends, and identifying future research frontiers such as knowledge preservation, reasoning interpretability, cross-lingual generalization, and multi-modal integration. It also provides a public resource hub at etods.net to accelerate community progress.

ABSTRACT

End-to-end task-oriented dialogue (EToD) can directly generate responses in an end-to-end fashion without modular training, which attracts escalating popularity. The advancement of deep neural networks, especially the successful use of large pre-trained models, has further led to significant progress in EToD research in recent years. In this paper, we present a thorough review and provide a unified perspective to summarize existing approaches as well as recent trends to advance the development of EToD research. The contributions of this paper can be summarized: (1) extbf{ extit{First survey}}: to our knowledge, we take the first step to present a thorough survey of this research field; (2) extbf{ extit{New taxonomy}}: we first introduce a unified perspective for EToD, including (i) extit{Modularly EToD} and (ii) extit{Fully EToD}; (3) extbf{ extit{New Frontiers}}: we discuss some potential frontier areas as well as the corresponding challenges, hoping to spur breakthrough research in EToD field; (4) extbf{ extit{Abundant resources}}: we build a public website\footnote{We collect the related papers, baseline projects, and leaderboards for the community at \url{https://etods.net/}.}, where EToD researchers could directly access the recent progress. We hope this work can serve as a thorough reference for the EToD research community.

Motivation & Objective

• To provide the first systematic and thorough survey of end-to-end task-oriented dialogue (EToD) research, addressing the lack of a unified overview in the field.
• To introduce a novel taxonomy distinguishing between modularly EToD (with intermediate supervision) and fully EToD (with end-to-end sequence-to-sequence training and differentiable KB retrieval).
• To identify and analyze emerging frontiers in EToD, including knowledge preservation, reasoning interpretability, zero-shot cross-lingual transfer, and multi-modal integration.
• To compile and curate a public repository of papers, baselines, and leaderboards at etods.net to support ongoing research and benchmarking in the EToD community.
• To guide future research by outlining key challenges and opportunities in knowledge grounding, model interpretability, and cross-lingual scalability.

Proposed method

• Proposes a unified taxonomy of EToD systems into two main categories: (1) modularly end-to-end, where components are trained jointly but retain some modular structure; and (2) fully end-to-end, where all components are trained end-to-end with a single sequence-to-sequence framework.
• Classifies existing EToD methods based on their knowledge representation mechanisms: entity triplet, row-level, and graph-based representations for KB integration.
• Reviews the use of large pre-trained models (PLMs) in modularly EToD to enhance performance and generalization, especially in few-shot and zero-shot settings.
• Analyzes differentiable attention mechanisms for KB retrieval in fully EToD, emphasizing their role in enabling end-to-end training and gradient backpropagation.
• Introduces and discusses the use of chain-of-thought reasoning in KB retrieval to improve interpretability and traceability of model decisions.
• Proposes the need for unified cross-lingual models that can align heterogeneous dialogue components (DST, DPL, NLG) across multiple languages, particularly for zero-shot transfer.

Experimental results

Research questions

• RQ1 How can end-to-end task-oriented dialogue systems be systematically categorized and unified under a single taxonomy?
• RQ2 What are the key differences and trade-offs between modularly and fully end-to-end EToD architectures in terms of performance, error propagation, and knowledge integration?
• RQ3 How can knowledge preservation be achieved when transferring from modularized, pre-trained dialogue models to end-to-end frameworks?
• RQ4 What challenges hinder the interpretability of KB retrieval in fully EToD models, and how can chain-of-thought reasoning improve transparency?
• RQ5 What are the main barriers to zero-shot cross-lingual EToD, particularly in knowledge base alignment and unified multi-lingual modeling?

Key findings

• The survey establishes a new taxonomy of EToD into modularly and fully end-to-end systems, providing a foundational framework for classifying existing and future methods.
• Modularly EToD methods with pre-trained models (PLMs) show strong performance, especially in few-shot and zero-shot settings, due to better generalization and transfer learning capabilities.
• Fully EToD models using differentiable attention for KB retrieval achieve state-of-the-art results but suffer from black-box behavior, limiting interpretability and trust.
• Chain-of-thought reasoning in KB retrieval is a promising direction to improve model transparency, though designing effective reasoning steps remains a non-trivial challenge.
• Cross-lingual EToD is constrained by knowledge base alignment and the lack of unified models for heterogeneous dialogue components across languages, with BiToD and GlobalWoZ emerging as key benchmarks.
• Multi-modal EToD is limited by small-scale and low-diversity benchmarks such as MMConv and SIMMC 2.0, highlighting the need for larger, more diverse multi-modal datasets to drive progress.