I can confirm this. I did NER and most tokens are not names entities, so they are "other". It's really hard to define what "other" means, even with lots of text the model is unsure. No matter how much "other" I provide, I couldn't train a negative class properly.