[Paper Review] Making Pre-trained Language Models Better Few-shot Learners

Tianyu Gao, Adam Fisch+1 more|arXiv (Cornell University)|Dec 31, 2020

Topic Modeling46 references128 citations

TL;DR

LM-BFF combines prompt-based fine-tuning, automatic prompt generation, and selective demonstrations to greatly improve few-shot learning with medium-sized language models, achieving up to 30% absolute gains over standard fine-tuning. It demonstrates strong task-agnostic performance across classification and regression with RoBERTa-large in few-shot settings.

ABSTRACT

The recent GPT-3 model (Brown et al., 2020) achieves remarkable few-shot performance solely by leveraging a natural-language prompt and a few task demonstrations as input context. Inspired by their findings, we study few-shot learning in a more practical scenario, where we use smaller language models for which fine-tuning is computationally efficient. We present LM-BFF--better few-shot fine-tuning of language models--a suite of simple and complementary techniques for fine-tuning language models on a small number of annotated examples. Our approach includes (1) prompt-based fine-tuning together with a novel pipeline for automating prompt generation; and (2) a refined strategy for dynamically and selectively incorporating demonstrations into each context. Finally, we present a systematic evaluation for analyzing few-shot performance on a range of NLP tasks, including classification and regression. Our experiments demonstrate that our methods combine to dramatically outperform standard fine-tuning procedures in this low resource setting, achieving up to 30% absolute improvement, and 11% on average across all tasks. Our approach makes minimal assumptions on task resources and domain expertise, and hence constitutes a strong task-agnostic method for few-shot learning.

Motivation & Objective

• Motivate practical few-shot learning with moderately-sized LMs (e.g., RoBERTa/BERT) instead of huge models like GPT-3.
• Develop a task-agnostic, simple set of techniques to improve fine-tuning with minimal data.
• Evaluate prompt-based fine-tuning and demonstration strategies across multiple NLP tasks (classification and regression).
• Provide an automated workflow for generating prompts and demonstrations to reduce manual engineering.

Proposed method

• Prompt-based fine-tuning that treats downstream tasks as masked language modeling with label words.
• Automatic prompt generation via (i) automatic selection of label words (verbalizers) and (ii) automatic generation of templates using a T5-based search.
• Dynamic and selective demonstrations where a single per-class example is sampled and paired with the input to form minimal demonstration sets.
• A systematic evaluation across 8 single-sentence and 7 sentence-pair tasks, using multiple random splits to assess stability.
• Comparison with standard fine-tuning and GPT-3-style in-context learning under few-shot settings.

Experimental results

Research questions

• RQ1 Can prompt-based fine-tuning with automatically generated prompts match or exceed manually designed prompts in few-shot regimes?
• RQ2 Does incorporating demonstrations with careful sampling improve performance beyond standard fine-tuning for medium-sized LMs?
• RQ3 How do automatic label-word selection and automated template generation contribute to robust few-shot learning across tasks (classification and regression)?
• RQ4 What is the impact of demonstrations sampling strategy and template quality on few-shot performance?
• RQ5 Is the approach task-agnostic and resource-efficient enough to be practical for real-world use?

Key findings

• Prompt-based fine-tuning substantially outperforms standard fine-tuning in few-shot settings.
• Automatic prompt generation (templates and label words) can match or outperform manual prompts across several tasks.
• Incorporating demonstrations with carefully designed sampling strategies yields additional gains in few-shot performance.
• The combined LM-BFF methods achieve up to 30% absolute improvement and 11% average improvement across evaluated tasks.
• On RoBERTa-large with 32 training examples, many binary SST-2-like tasks reach around 90% accuracy.