Closed GiovanniPioDelvecchio closed 1 year ago
paper mentioned in the dataset one. Allows learning rules capturing relations between labels: https://link.springer.com/article/10.1007/s10994-016-5552-1
the more suitable models for testing are: ELECTRA ConvBERT BORT Funnel Transformers the others could f**k themselves
BORT tested with the same head, gave 0.24 F1 i think it can be discarded
Funnel gave 0.3765 F1 but probably it compresses the model too much. Probably one could try to finetune it directly.
It is necessary to extend the baseline by adding a language model at its base, in particular one can choose between the following transformer-based language models: