Emotional Polar Predictions Based on Matching Prompts
Zhangping Yang, Ye Xia, Hantao Xu, Pujie Zhao · 2023
In recent years, pre-trained language models have motivated the study in various tasks in natural language processing(NLP). In a new paradigm instead of fine-tuning, reformulating downstream tasks as cloze tasks by add textual or continuous prompts is a popular measure for solving NLP tasks. However, the effect of prompt learning tends to be only ensured by appropriate prompts and precise verbalizers. To address this issue, we propose a prompt matching method based on trigger tokens to solve emotional polar prediction tasks. The proposed model can extract the aspect words and key adjectives from input text as trigger tokens. Meanwhile, by introducing external knowledge we expand trigger tokens which are applicated in verbalizers and prompts matching. According to experiments, we show that providing instances which share the similar emotional trigger tokens improves model performance. And tests on the hotel reviews dataset show that our model have more high precision rate. In few-shot scenario, We also show that the fewer the samples, the more our model outperforms fine-tuning models. These results demonstrate that automatically providing instances with similar emotions is a new method to solve text classification tasks in few-shot scenario.