What I learned analysing Kaggle Deep Past Challenge

Samenvatting

I fell into a rabbit hole looking at Kaggle’s Deep Past Challenge and ended up reading a bunch of winning solution writeups. Here's what I learned At first glance it looks like a machine translation competition: translate Old Assyrian transliterations into English. But after reading the top solutions, I don’t think that’s really what it was. It was more like a data construction / data cleaning competition with a translation model at the end. Why: the official train set was tiny: 1,561 pairs t...