search

n-waves / ulmfit4de

ULMFiT Method for German Language
15 stars 1 forks source link

Expanding Emoji to German words #5

Open PiotrCzapla opened 6 years ago

PiotrCzapla commented 6 years ago

as discussed on the forum I would love to see how this improves the model.

MicPie commented 6 years ago

Dear Piotr,

thank you for the organization of the ulmfit4de repo!

I am currently trying to get the german emoji description with beautiful soup into python and then into a csv.

I will get back to you when I have them and checked them.

Best regards Michael

MicPie commented 6 years ago

Dear Piotr,

see the link posted to the xml files with the different emoji translation by Marcin on the fastai-forum: http://forums.fast.ai/t/ulmfit-german/22529/49

Best regards Michael

PiotrCzapla commented 5 years ago

Thx. Have you had some time to give it a try?

MicPie commented 5 years ago

Dear Piotr,

sorry, for my late reply. - I was busy with get going with the fastai v1 library.

I am also following the ulmfit threads on the fastai forums from you. What is you plan with the upcoming fastai text v1 and Google BERT?

I will try to cut off some time on weekend to look into the ulmfit4de repo.

Best regards Michael

PiotrCzapla commented 5 years ago

Don't worry following closely fastaiv1 and doing all the projects will get you a long way so it is a good choice. Re. BERT we want to see how much better /worst it is from ULMFiT on classification problems.
ULMFiT is much faster and easier to train than BERT so I think it still has some good use cases. But we need to compare to see what we are losing / gaining.

btw. I think DE is pretty much done we exceeded SOTA for GE17 and we are very close to STOA GE18, the only thing is to consolidate languages in fastaiv1, and maybe try to use biLM training but that is for later.