0

I often see variants of IOB tagging scheme such as IOB, BIO, IOBES mentioned in the literature for chunking, NER etc. I tried using only BI tags for detecting morpheme boundaries (segmentation) in a binary classification setting and got high F1 score while adding O tag gave much lower score. Is there any drawback with using only BI tags ? Why I am not seeing people using it ?

math
  • 49
  • 1
  • 7

1 Answers1

0

BI or IO are easy to implement but should be worse than IOBES for NER tasks.

Have you tried using other tagging schemes? Intuitively, more elaborate schemes are better suited for longer sequences. For NER, BIO or IOBES are reported to reach significantly better F1 scores or a good rule-of-thumb. You can also try BILOU or BIL2 which showed good results on SOV languages (Japanese, Korean, Urdu).

minerals
  • 6,090
  • 17
  • 62
  • 107