[Paper Review] SOLOIST: Few-shot Task-Oriented Dialog with A Single Pre-trained Auto-regressive Model

Baolin Peng, Chunyuan Li+4 more|arXiv (Cornell University)|May 11, 2020

Topic ModelingComputer Science42 references88 citations

TL;DR

SOLOIST proposes a single pre-trained auto-regressive Transformer model that unifies dialog state tracking, policy learning, and response generation for task-oriented dialog. By leveraging transfer learning and machine teaching, it achieves state-of-the-art performance on CamRest and MultiWOZ with only a few task-specific examples, significantly reducing labeling costs.

ABSTRACT

This paper presents a new method SOLOIST, which uses transfer learning to efficiently build task-oriented dialog systems at scale. We parameterize a dialog system using a Transformer-based auto-regressive language model, which subsumes different dialog modules (e.g., state tracker, dialog policy, response generator) into a single neural model. We pre-train, on large heterogeneous dialog corpora, a large-scale Transformer model which can generate dialog responses grounded in user goals and real-world knowledge for task completion. The pre-trained model can be efficiently adapted to accomplish a new dialog task with a handful of task-specific dialogs via machine teaching. Our experiments demonstrate that (i) SOLOIST creates new state-of-the-art results on two well-known benchmarks, CamRest and MultiWOZ, (ii) in the few-shot learning setting, the dialog systems developed by SOLOIST significantly outperform those developed by existing methods, and (iii) the use of machine teaching substantially reduces the labeling cost. We will release our code and pre-trained models for reproducible research.

Motivation & Objective

• To develop a scalable, unified framework for task-oriented dialog systems that reduces dependency on task-specific module design.
• To minimize labeling costs in few-shot dialog learning by leveraging machine teaching on a pre-trained model.
• To improve zero-shot and few-shot generalization in dialog systems using transfer learning from large-scale heterogeneous dialog corpora.
• To unify state tracking, policy learning, and response generation into a single end-to-end auto-regressive model.
• To achieve state-of-the-art performance on benchmark datasets like CamRest and MultiWOZ with minimal fine-tuning data.

Proposed method

• Parameterize the entire dialog system using a single Transformer-based auto-regressive language model, replacing separate modules for state tracking, policy, and response generation.
• Pre-train the model on large-scale, heterogeneous dialog corpora to learn grounded responses and user goal understanding.
• Adapt the pre-trained model to new tasks using machine teaching, where only a few task-specific dialog examples are used to guide fine-tuning.
• Use a single sequence-to-sequence autoregressive generation head to produce responses conditioned on user goals and real-world knowledge.
• Integrate user goals and external knowledge explicitly into the input prompt to ensure grounded and task-compliant responses.
• Leverage transfer learning to enable rapid adaptation to new domains with minimal labeled data.

Experimental results

Research questions

• RQ1 Can a single pre-trained auto-regressive model effectively unify state tracking, policy learning, and response generation in task-oriented dialog?
• RQ2 How does SOLOIST perform in few-shot settings compared to existing methods on standard benchmarks?
• RQ3 To what extent does machine teaching reduce the labeling cost while maintaining or improving model performance?
• RQ4 Can transfer learning from large-scale heterogeneous dialog data improve generalization in low-resource dialog scenarios?
• RQ5 Does the unified architecture of SOLOIST achieve state-of-the-art results on CamRest and MultiWOZ under few-shot conditions?

Key findings

• SOLOIST achieves new state-of-the-art results on both CamRest and MultiWOZ benchmarks in the few-shot setting.
• The model significantly outperforms existing methods in few-shot learning, demonstrating superior generalization with minimal labeled data.
• Machine teaching reduces the required labeling cost for new dialog tasks while maintaining high performance.
• The unified auto-regressive model effectively captures complex dialog dynamics without separate module training.
• Pre-training on large-scale heterogeneous dialog corpora enables strong zero-shot and few-shot adaptation capabilities.
• The model generates grounded, task-compliant responses by conditioning on user goals and real-world knowledge.