[Paper Review] SOLOIST: Building Task Bots at Scale with Transfer Learning and Machine Teaching

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

Topic Modeling56 references99 citations

TL;DR

SOLOIST builds task-oriented dialogue bots with a single Transformer model pre-trained on diverse dialog data and fine-tuned via machine teaching, achieving state-of-the-art results and reduced labeling costs in few-shot scenarios.

ABSTRACT

We present a new method SOLOIST that uses transfer learning and machine teaching to build task bots at scale. We parameterize classical modular task-oriented dialog systems using a Transformer-based auto-regressive language model, which subsumes different dialog modules into a single neural model. We pre-train, on heterogeneous dialog corpora, a task-grounded response generation 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 new tasks with a handful of task-specific dialogs via machine teaching, where training samples are generated by human teachers interacting with the system. Experiments show that (i) SOLOIST creates new state-of-the-art on well-studied task-oriented dialog benchmarks, including CamRest676 and MultiWOZ; (ii) in the few-shot fine-tuning settings, SOLOIST significantly outperforms existing methods, and (iii) the use of machine teaching substantially reduces the labeling cost of fine-tuning. The pre-trained models and codes are available at https://aka.ms/soloist.

Motivation & Objective

• Eases the workflow for building many task bots by using a single, pre-trained model.
• Unifies NLU, DST, POL, and NLG into one transformer-based architecture.
• Leverages task-grounded pre-training on heterogeneous dialog corpora.
• Fine-tunes to new tasks with minimal task-specific data via machine teaching.
• Demonstrates state-of-the-art performance and reduced labeling cost in few-shot settings.

Proposed method

• Parameterizes modular task-oriented dialog as a Transformer-based autoregressive model.
• Pre-trains a task-grounded response generation model on heterogeneous dialog data, learning belief tracking and grounded responses.
• Decomposes the training objective into belief prediction and grounded response generation with a contrastive objective.
• Implements a multi-task pre-training objective L_theta combining L_B, L_R, and L_C.
• Fine-tunes on new tasks with few dialogs; applies machine teaching to generate corrective training samples.
• Uses nucleus sampling for decoding and larger model variants to enhance few-shot learning.

Experimental results

Research questions

• RQ1 How does SOLOIST perform on standard task-oriented dialog benchmarks compared to state-of-the-art methods?
• RQ2 Can SOLOIST generalize effectively to new domains in few-shot fine-tuning settings?
• RQ3 How effective is machine teaching for fine-tuning task bots and reducing labeling cost?
• RQ4 What is the impact of model size on few-shot transfer performance?
• RQ5 What is the role of task-grounded pre-training in enabling end-to-end grounding and state tracking?

Key findings

• SOLOIST achieves state-of-the-art results on CamRest676 and MultiWOZ benchmarks.
• In few-shot fine-tuning settings, SOLOIST significantly outperforms existing methods.
• Machine teaching substantially reduces labeling cost during fine-tuning.
• A larger model (SOLOIST-L) consistently outperforms the base model in few-shot scenarios.
• SOLOIST maintains strong performance on component tasks like intent classification, slot filling, and dialog state tracking.
• Compared to baselines, SOLOIST requires fewer domain-specific annotations while delivering strong end-to-end performance.