[Paper Review] PPT: Pre-trained Prompt Tuning for Few-shot Learning

Yuxian Gu, Xu Han+2 more|arXiv (Cornell University)|Sep 9, 2021

Topic Modeling100 citations

TL;DR

PPT pre-trains soft prompts on unified self-supervised tasks to initialize prompt tuning, enabling PPT to outperform vanilla prompt tuning and often rival or exceed full-model fine-tuning in few-shot and full-data settings.

ABSTRACT

Prompts for pre-trained language models (PLMs) have shown remarkable performance by bridging the gap between pre-training tasks and various downstream tasks. Among these methods, prompt tuning, which freezes PLMs and only tunes soft prompts, provides an efficient and effective solution for adapting large-scale PLMs to downstream tasks. However, prompt tuning is yet to be fully explored. In our pilot experiments, we find that prompt tuning performs comparably with conventional full-model fine-tuning when downstream data are sufficient, whereas it performs much worse under few-shot learning settings, which may hinder the application of prompt tuning in practice. We attribute this low performance to the manner of initializing soft prompts. Therefore, in this work, we propose to pre-train prompts by adding soft prompts into the pre-training stage to obtain a better initialization. We name this Pre-trained Prompt Tuning framework "PPT". To ensure the generalization of PPT, we formulate similar classification tasks into a unified task form and pre-train soft prompts for this unified task. Extensive experiments show that tuning pre-trained prompts for downstream tasks can reach or even outperform full-model fine-tuning under both full-data and few-shot settings. Our approach is effective and efficient for using large-scale PLMs in practice.

Motivation & Objective

• Motivate the need to bridge pre-training with downstream tasks via prompt tuning for large PLMs.
• Propose a pre-training strategy for soft prompts to improve initialization under few-shot settings.
• Unify downstream classification tasks into a common pre-training framework to generalize prompts across tasks.
• Demonstrate that PPT can achieve or surpass full-model fine-tuning while maintaining parameter efficiency.

Proposed method

• Represent downstream tasks as pattern-verbalizer pairs to formulate prompts.
• Pre-train soft prompts on self-supervised tasks aligned with task formats (sentence-pair, multiple-choice, single-text).
• Unify tasks into a single multiple-choice pre-training format to enable broad applicability.
• Initialize downstream PT with the pre-trained soft prompts and fine-tune only 0.41M prompt parameters.
• Evaluate PPT and variants on English and Chinese 11B PLMs across few-shot and full-data settings.

Experimental results

Research questions

• RQ1 Can pre-trained soft prompts improve the effectiveness of prompt tuning in few-shot learning for large PLMs?
• RQ2 Does unifying task formats via pre-training enhance cross-task generalization of prompts?
• RQ3 How does PPT compare to full-model fine-tuning and vanilla prompt tuning in terms of accuracy and variance under few-shot and full-data regimes?

Key findings

• PPT generally outperforms vanilla prompt tuning and language-model adaptation baselines in few-shot and full-data scenarios.
• Hybrid PPT (soft prompts with carefully designed hard prompts) often yields the best performance on several English and Chinese tasks.
• PPT can surpass or closely approach full-model fine-tuning (FT) on many datasets, indicating a gap between pre-training and downstream tasks is bridged by pre-training prompts.
• Unified PPT (unifying task formats into multiple-choice) achieves competitive results, especially for tasks with more than five labels.
• PPT reduces variability in few-shot results, yielding more stable performance across seeds.