Errors made by neural models

Interrogative adjectives

It may seem strange but gani is treated as an interrogative adjective in the Helsinki corpus of Swahili and by Mohammed (2001). This is probably due to analogy with -pi and -ngapi which are both inflected with adjectival concord.

Hashtags need to be rejoined

The tokenizer used split pound signs from the rest of the hash tag: #NairobiBlast -> # NairobiBlast.

These should all be rejoined.


-ote meaning 'all' is nearly always labeled as adv instead of det like it should be.