Support the case where the model can predict NER entities but doesn't in that case
Seen on a model that was trained both at:
- object-level (no NER entities)
- page-level (one NER entity per object).
When running on objects, it doesn't predict any entities. If the user provides the tokens used during training, the prediction fails. See https://arkindex.teklia.com/process/16be553d-a392-4334-b446-58cec51d758b/0
Edited by Yoann Schneider