token_type_id is converted into the segment embedding. For question answering, this needs to highlight whether a token belongs to sequence 0 or 1. encode_plus takes care of correctly setting this parameter automatically.
