As stated earlier, the encoder will process the source input sequence through an LSTM and encode the source text into a meaningful summary. The meaningful summary would be stored in the final sequence step hidden and cell states hf and cf. These vectors together (that is, [hf; cf]) provide a meaningful context about the source text, and the decoder is trained to produce its own target sequence conditioned on the hidden and cell state vectors [hf; cf].
Illustrated in the following diagram, Figure 3.5, is a detailed diagram of the training process of an English-to-French translation. The English sentence It's a beautiful day is converted to a meaning summary through an LSTM, which is then stored ...
