In neural machine translation, a bidirectional encoder processes the source sequence in both directions, capturing forward and backward context. This lets the model learn long-range alignments without reversing the input. section 5.7
Reversal versus two-direction encoding
Reversing the source improved unidirectional RNN encoders, but empirical comparisons found that bidirectional encoders outperformed unidirectional encoders even with reversal. The reversal advantage also declined as encoder depth increased. section 5.7
