What means accuracy in the results of the training of a seq2seq model?

Question

I trained a machine translation model using recurrent networks in a seq2seq model. I get 48% of accuracy, what is the meaning of this??

Answer 1

At training time, seq2seq is approach as a sequence labeling problem. You just assign symbols labels of what the next symbol is. From this perspective, you can also measure accuracy, simply as a proportion of cases when a correct label is assigned.

However, it tells nothing about the performance at inference time. At inference time, the seq2seq model does not use the ground truth labels, but its own outputs from the previous step. The proper validation could be done only at the entirely generated sequence using a sequence-level metric such as BLEU score.