Korean Neural Machine Translation Using Hierarchical Word Structure

Korean neural machine translation may significantly suffer from low-resource issues. We thus propose an enhancement method that fully exploits the hierarchical Korean word embedding structure from source representation. To our best knowledge, this is the first attempt for such Korean NMT tasks. Every Korean word can be decomposed into character- and jamo-level (sub-character) units. Therefore, We merge the character- and jamo-level representations with word embeddings to capture important combining word meaning. And then the merged representations are fed into NMT model. Our simple and novel method achieves BLEU improvements (up to 0.6) compared to word-based NMT baselines on Korean-to-Chinese and Korean-to-English translation tasks.