Integrated-gradient word importance scores outperform attention and black-box baselines for predicting which input words affect BLEU and which words are under-translated.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2019 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Towards Understanding Neural Machine Translation with Word Importance
Integrated-gradient word importance scores outperform attention and black-box baselines for predicting which input words affect BLEU and which words are under-translated.