Ner_crf recognizes hypen "-" entity that does not exist in training data

Hey guys, i’m experiencing some weird behavior from the ner_crf component. It recognizes some of my entities as a “hypen entity” (“entity”: “-”). Has anybody else had this issue?

"entities": [
                {
                    "confidence": 0.6524294862065713,
                    "end": 6,
                    "entity": "-",
                    "extractor": "ner_crf",
                    "start": 0,
                    "value": "burger"
                }
            ]

There is no training examples with an “-” entity.

Rasa NLU version 0.13.0

Are you sure you haven’t accidentally marked an entity like that? What format is your training data in?