Authors: jianqiang liu, kai zhou, chen duan, shuaishuai chang, ran wei, shan guo

Affiliation: Meituan

Description: We are Meituan-OCR team. Our method follows the framework of LiLT and designs a jointly training scheme in which SER and RE task are optimized jointly, benefiting greatly for task SER and RE respectively. We only use SVRD-2023 dataset to train the above-mentioned model. In the post-process stage, we design some rules to merge the keys pointing to the same value. Specifically, the table recognition result is served as an alternative information to handle the table format page. Besides, we design an adaptive-inference scheme to tackle the situation in which the relations between long-range keys and values are missing because they are in different 512-token inference items.