Attention and bidirectional contextThe neural classifier's encoder is a 6-layer transformer with bidirectional