Hi! If you want to use binary annotation with a model in the loop, you’re always giving feedback on the suggestion in this exact context. So you should definitely reject incomplete spans. This way, you’re telling the model “no, this particular analysis of the text is incorrect”, the weights will be updated to reflect that particular decision and the model will “try again” with a different analysis, hopefully moving towards more correct entity boundaries.
That said, if your data contains a lot of fairly abstract multi-token entities like that and the model struggles, it might take pretty long until it converges (or it might not converge at all). You could try adding some --patterns, or collect a small set with ner.manual or ner.make-gold that covers the especially complex entities, pre-train the model with that and then improve that pre-trained model further with ner.teach. You might also want to check out this thread, which discusses an approach for extending entity boundaries with rules: Expanding NER to include neighbouring tokens