Congratulations to all the winners.

Sorry for the delay in posting the discussion. However, I hope this will be helpful in the future.

It is made up of two BERT models. (bert-base-uncaed, bert-base-arabertv02)

I used the Arabizi-Arabic translation. Thanks for the module.

https://zindi.africa/competitions/ai4d-icompass-social-media-sentiment-analysis-for-tunisian-arabizi/discussions/5256

I also added traditional models (LSTM, fasttex, etc...) to the ensemble to boost the score.

The detailed solution is available on github.

The final sub is 0.804, but if we choose the right sub it is 0.842

https://github.com/ktrw1011/sentiment-analysis-for-tunisian-arabizi