Enhancing Cross-lingual Prompting with Mask Token Augmentation

Prompting shows promising results in few-shot scenarios. However, its strength for multilingual/cross-lingual problems has not been fully exploited. Zhao and Sch\"utze (2021) made initial explorations in this direction by presenting that cross-lingual prompting outperforms cross-lingual finetuning. In this paper, we conduct empirical analysis on the effect of each component in cross-lingual prompting and derive Universal Prompting across languages, which helps alleviate the discrepancies between source-language training and target-language inference. Based on this, we propose a mask token augmentation framework to further improve the performance of prompt-based cross-lingual transfer. Notably, for XNLI, our method achieves 46.54% with only 16 English training examples per class, significantly better than 34.99% of finetuning.