Encoder hidden state (hᵢ)

The vector produced by the bidirectional encoder for source position i. Encodes information about word i in the context of the full sentence. Bahdanau's model retains all T encoder hidden states (one per source word) throughout decoding, rather than discarding them after encoding.