Character-Aware Decoder for Neural Machine Translation

[zomux]

Main Authors / Organization

Adithya Renduchintala∗ and Pamela Shapiro∗ and Kevin Duh and Philipp Koehn

Johns Hopkins University

PDF link

https://arxiv.org/abs/1809.02223

Hypothesis

Making embeddings character-aware helps capturing mophology even for subwords

Approach

Using a gate function to combine the conventional word embeddings and composed character embeddings

Main Experimental Result

De-En task + 1.2 BLEU in low-resource setting