The paper introduces a method called Label-Aware Automatic Verbalizer (LAA V) to enhance the effectiveness of few-shot text classification. This technique augments manual labels to achieve better classification results. The verbalizer translates output from a language model into a predicted class. The experimental results across five datasets in five languages show that LAA V significantly outperforms existing verbalizers, especially in mid-to-low resource languages. It suggests more relevant words compared to similar approaches.
Publication date: 19 Oct 2023
Project Page: https://arxiv.org/abs/2310.12778
Paper: https://arxiv.org/pdf/2310.12778
