Based on a record of dissents on FOMC votes and transcripts of the meetings from 1976 to 2017, we develop a deep learning model based on self-attention mechanism to create a measure of disagreement for members in each meeting. While dissents are rare, we find that members often have reservations wit...