Generating Training Data with Language Models: Towards Zero-Shot Language Understanding

TL;DR

This paper introduces a zero-shot learning approach for natural language understanding by generating training data with language models, achieving competitive results on the GLUE benchmark without task-specific data.

Contribution

The authors propose a novel method combining unidirectional and bidirectional PLMs to generate training data for zero-shot NLU, improving performance over existing prompting methods.

Findings

Achieves strong zero-shot performance on GLUE tasks

Outperforms zero-shot prompting methods significantly

Comparable to few-shot approaches with limited data

Abstract

Pretrained language models (PLMs) have demonstrated remarkable performance in various natural language processing tasks: Unidirectional PLMs (e.g., GPT) are well known for their superior text generation capabilities; bidirectional PLMs (e.g., BERT) have been the prominent choice for natural language understanding (NLU) tasks. While both types of models have achieved promising few-shot learning performance, their potential for zero-shot learning has been underexplored. In this paper, we present a simple approach that uses both types of PLMs for fully zero-shot learning of NLU tasks without requiring any task-specific data: A unidirectional PLM generates class-conditioned texts guided by prompts, which are used as the training data for fine-tuning a bidirectional PLM. With quality training data selected based on the generation probability and regularization techniques (label smoothing and temporal ensembling) applied to the fine-tuning stage for better generalization and stability, our approach demonstrates strong performance across seven classification tasks of the GLUE benchmark (e.g., 72.3/73.8 on MNLI-m/mm and 92.8 on SST-2), significantly outperforming zero-shot prompting methods and achieving even comparable results to strong few-shot approaches using 32 training samples per class.