NER and Relation Extraction

[ljvmiranda921 (Lj Miranda)]

Hi @koayst ,

Welcome to Prodigy

@koayst

• I annotate the dataset first by labelling the entities and then build a NER model
• Use the NER model generated and run it on the same dataset as in step #1 and together with the entities inferred by the model, label the relation between entities and build another relation model.

For the entity relation extraction (REL) use-case, your first option of
annotating them separately makes sense. This is because for most REL applications, deciding whether something is an entity is unrelated to whether that
entity is a relation, so you have different thought processes in annotating.

If you annotate them at the same time, you might miss out on actual entities that just happened to not belong in a relation--lowering your recall and affect the consistency of the NER data and subsequent models.

I also suggest looking at these discussions for more information:

• Information extraction from legislative text - Doubts and questions
• Can Dependencies & Relations work after Span Categorization?