The strategy that we'll adopt to build a machine translation system using the encoder decoder architecture is as follows:
- We have two decoder datasets while preparing input and output datasets:
- The decoder_input_data combined with encoder_input_data is the input and decoder_target_data is the output
- The decoder_input_data starts with the start word
- When we are predicting the first word in the decoder, we are using the input set of words, converting them into a vector, which then gets passed through a decoder model that has start as input. The expected output is the first word after start in output
- We proceed in a similar manner, where the actual first word in the output is the input, while predicting the second word
- We'll calculate the accuracy of model based on this strategy