Hey,
What is the bias feature in NER_CRF? It seems that is impacting my results negatively. With the bias feature I have a lot of balancing feature weights for the O and entity label. E.g. -1.11 for the entity label and 1.11 for the O-label. So one label is almost the same but sign inverted. In this cases It seems the CRF randomly picks one of the label to be the entity. So in the above case it chooses the word to be an entity although it has a negative sign on its weights.
I feel I miss something about the weights of the features?