Implementation details for baseline results

[Yusia7]

Hi ,

Thanks for sharing this great work! I've been studying your paper and have a question about the baseline results shown in Table 3.
I was wondering if you could kindly clarify which specific implementation you used to generate these baseline results? If it's possible, would you be willing to share the relevant code used for these experiments?
This would be extremely helpful for reproducing the results.

Thanks in advance!

[shirley-wu]

Hi, sorry for the late response. Are you referring to the TAPEX and TAPAS baseline? For TAPAS, since it's already pretrained to perform table QA task, we directly use the pretrained model. We provide the user query as input, and run the model to select important cells that can help answer the input question. Then, we prompt ChatGPT to interpret the selected cells into the output text. For TAPEX, since it's not pretrained on table QA task, we finetune the model on the weak supervision training set, then use it to generate the output. It's explained in Section 5.1. Do you need more details or clarification on anything?