Finetune/train a model for RTL scripts/languages

[johnlockejrr]

Congratulations for this marvelous idea and project!

Could I finetune a DANIEL model for RTL scripts like Hebrew, Arabic, Samaritan etc.?
If so, would that work out of the box or should I implemment other steps (python-bidi etc.)?

Thank you!

[Shulk97]

Thank you for your interest in my work!

DANIEL is perfectly capable of processing right-to-left languages, but it is likely that this will require a larger than usual amount of annotated data or to design new synthetic data including the right-to-left reading order.
Indeed, DANIEL has never encountered this reading order and will therefore require an adaptation phase during training to modify its attention.
If the documents you wish to process contain only right-to-left text, perhaps flipping the image on the vertical axis might do the work.