What do the flat_ner and span_mode parameters do?

[omri374]

I'm trying to better understand how these two parameters affect the output. Any intuition would be great, especially around recall vs. precision of entity recognition, i.e. how does choosing these parameters influence the types and spans of entities detected.

[urchade]

Hello @omri374

• span_mode determines how spans are represented and does not change after training, I don't think it influence much recall/precision
• flat_ner enables or disables nested entities (set flat_ner=False for nested entities).
• multi_label allows a single span to have multiple entity types.

To increase recall, you can set flat_ner=False and decrease the threshold.

Let me know if you have any further questions!

[omri374]

Thanks @urchade, could you give an example of a nested entity compared to a flat one?

[urchade]

@omri374 flat_ner (a) does not allow overlapping entities, while nested (b) does (flat_ner=False)

[omri374]

Thanks