Prompting ELECTRA: Few-Shot Learning with Discriminative Pre-Trained Models
This work addresses the challenge of enabling efficient few-shot learning for discriminative models, offering a method that improves performance without adding parameters, which is incremental but practical for NLP applications.
The paper tackled the problem of adapting prompt-based few-shot learning to discriminative pre-trained models like ELECTRA, which traditionally don't fit the text infilling paradigm used by masked language models, and showed that ELECTRA outperforms masked language models across a wide range of tasks.
Pre-trained masked language models successfully perform few-shot learning by formulating downstream tasks as text infilling. However, as a strong alternative in full-shot settings, discriminative pre-trained models like ELECTRA do not fit into the paradigm. In this work, we adapt prompt-based few-shot learning to ELECTRA and show that it outperforms masked language models in a wide range of tasks. ELECTRA is pre-trained to distinguish if a token is generated or original. We naturally extend that to prompt-based few-shot learning by training to score the originality of the target options without introducing new parameters. Our method can be easily adapted to tasks involving multi-token predictions without extra computation overhead. Analysis shows that ELECTRA learns distributions that align better with downstream tasks.