Authors: Kai Yang, Tingmao Lin, Ye Wang, Shuqiang Lin, Jian Xie, Bin Wang, Wentao Liu, Xiaolu Ding, Jun Zhu, Hongyan Pan, Jia Lv

Affiliation: CCB Financial Technology Co. Ltd, China

Description: Our method, the prompt learning based on UIE-X, which is a universal information extraction engine, make good prompt design on few-shot struct text extraction. By analyzing the data, it shows that the visual structured information, and the pure text features on it has no excessive continuous semantic information. We apply the prompt methodology based on the multimodal pre-trained model. And the prompt designed from label.json has show the semantic and position information of the entities. In order to optimize, we make a few negative examples for training task, and experiments show, only the few training examlpes has achieve a good result.