In the previous round of training, all the "empty" roles are filler tokens are annotated as O and treated as garbage strings, but now that the "linker" code is updated and is capable of handling empty r/f, the round 1 annotation need to be re-annotated (for O tokens) to match the new model behavior.
In the previous round of training, all the "empty" roles are filler tokens are annotated as
Oand treated as garbage strings, but now that the "linker" code is updated and is capable of handling empty r/f, the round 1 annotation need to be re-annotated (forOtokens) to match the new model behavior.