Few-Shot Text Generation with Pattern-Exploiting Training

TL;DR

This paper introduces GenPET, a novel method for few-shot text generation that leverages pattern-exploiting training with natural language instructions, improving performance on summarization and headline tasks.

Contribution

It extends pattern-exploiting training to generative tasks, addressing key challenges in task description clarity and overfitting, and demonstrates improved results in few-shot text generation.

Findings

GenPET outperforms strong baselines in few-shot summarization tasks.

Effective use of natural language instructions enhances generation quality.

The method improves data efficiency in text generation scenarios.

Abstract

Providing pretrained language models with simple task descriptions in natural language enables them to solve some tasks in a fully unsupervised fashion. Moreover, when combined with regular learning from examples, this idea yields impressive few-shot results for a wide range of text classification tasks. It is also a promising direction to improve data efficiency in generative settings, but there are several challenges to using a combination of task descriptions and example-based learning for text generation. In particular, it is crucial to find task descriptions that are easy to understand for the pretrained model and to ensure that it actually makes good use of them; furthermore, effective measures against overfitting have to be implemented. In this paper, we show how these challenges can be tackled: We introduce GenPET, a method for text generation that is based on pattern-exploiting training, a recent approach for combining textual instructions with supervised learning that only works for classification tasks. On several summarization and headline generation datasets, GenPET gives consistent improvements over strong baselines in few-shot settings.