Attention Weights in Transformer NMT Fail Aligning Words Between Sequences but Largely Explain Model Predictions.

Javier Ferrando, Marta R. Costa-juss

Browse the full EMNLP paper archive.