improve normalisation logic?

[derhuerst]

Thanks for this project, I think it is very underrated!

I came across a use case where I assume the normalisation doesn't work as intended:

1. Angermünde, Rosenstr. 53.01735, 14.00092
2. Rosenstrasse, Angermünde 53.01691, 14.00058

If I rename the second one to Rosenstraße, Angermünde, the two are successfully classified as similar.

[patrickbr]

Could you share your exact setup / command, so that I can reproduce this issue? IIRC, normalization is not enabled per default. It might actually be the case that the model never had the opportunity to learn the equivalence of "str." and "strasse".

[derhuerst]

Unfortunately, I don't know anymore how I had set things up.

I assume I was looking into how statsimi works with stops from the VBB GTFS dataset (mirror with old versions).