No, you can successfully fit complex models if you use byte-pair or similar encodings (morphessor [1] comes to mind). [1] https://morfessor.readthedocs.io/en/l

No, you can successfully fit complex models if you use byte-pair or similar encodings (morphessor [1] comes to mind). [1] https://morfessor.readthedocs.io/en/latest/ https://morfessor.readthedocs.io/en/latest/ You also will get much more meaningful embeddings from summing embeddings of part of the word.