About Adversarial Matching Details

[bonny...@gmail.com]

Hello, thank you for the great work.

I think your adversarial matching is great and want to reproduce it.

I have some questions about your adversarial matching.

1. In Aligning Detections, 

    Could you tell me details about the rule-based system? (e.g. 15% rj -> qi, 15% rj -> ri, 70% rj -> oi)

2. In Semantic categories,

    I understood dividing bucket as: 

      1) Divide into buckets according to the gender of the  pronoun in response

      2) In the buckets divided above, divide each of them again according to question types

   Do I understand it correctly? 

Thank you.

[Rowan Zellers]

hi! Sorry for the delay. Here's info that could hopefully help:

1. It's a bit more complicated, we looked at things on a case-by-case basis involving detections that were mentioned in one image/answer pair and not in another, etc. The remapping that I used is in the file called 'dataloader.py' under _remap_answer. However, I'm not sure that it matters too much, probably the most important thing is that you generate enough ways of realigning the sentences for a resulting text-only model to be somewhat confused.

2. Yep! exactly.

lmk if that helps, or if I can provide more info!

[이수련]

Hello,

Your reply and code are very helpful.

Thank you.

2020년 11월 4일 수요일 오전 4시 10분 4초 UTC+9에 Rowan Zellers님이 작성: