ABSTRACT Event Detection (ED) involves the recognition of trigger words and the classification of events. Existing methods of ED heavily depend on supervised learning that requires extensive labeled datasets. However, manual labeling is very expensive in practice. In this paper, we propose a novel approach by Fourier Transform Filter and Prompt Learning (i.e., FTF‐PL) to solve the issue of low resources in ED. We utilize a Fourier Transform Filter to process the word embedding vectors output from the pretrained model and filter the noise or redundant information in the word embedding vectors. Such a transformation enables the final vectors to be more aligned with the core semantics of the sentence. We also adopt prompt learning to obtain event‐related knowledge from pretrained language models. Extensive experimental results on the FewEvent and FewEvent++ benchmark datasets demonstrate that our proposed method can achieve better results compared to previous baselines, including PA‐CRF, HCL‐TAT, Meta‐Event, MPC‐CA, and LLaMA3.1. Furthermore, statistical significance analysis on FewEvent and FewEvent++ confirm that the improvements over the strongest baseline are statistically significant under different settings.
Xu et al. (Mon,) studied this question.