P-tuning has demonstrated that anchor tokens are beneficial for improving the performance of downstream tasks. However, manual selection of anchor tokens may result in subjective or suboptimal results. In this paper, we present aCat to automatically select anchor tokens. Following the framework of the soft-hard prompt paradigm, aCat achieves automatic prompt template construction. Experiments conducted on natural language understanding benchmarks demonstrate the effectiveness of our proposed method. On the seven SuperGLUE datasets, aCat achieves higher accuracy than P-tuning, with an average accuracy surpassing P-tuning V2.
Ye et al. (Tue,) studied this question.