The difference of results on the SER of FUNSD between the Table 2 and the Table 6 in the paper.

[WenjinW]

Hi! I like the idea of decoupling text and layout information to leverage existing pre-trained language models. I had some confusion when I was reading the paper.

• Table 2 shows that the F1 of LiLT[InfoXLM] on the SER task of FUNSD is 0.8586.
• However, in Table 6, the F1 is 0.8415.

Why are the performances reported in the two tables different?

Thanks for your reply.

[leitouran]

Hi, I had the same confusion but I think the results you mention for Table 2 are for the multilingual model, while the ones in Table 3 are English only.

[jpWang]

Hi, @WenjinW @leitouran , Table 2 follows the fine-tune style of https://github.com/microsoft/unilm/blob/master/layoutlmft/examples/run_funsd.py, but Table 3 follows the fine-tune style of https://github.com/microsoft/unilm/blob/master/layoutlmft/examples/run_xfun_ser.py.