Named entity transliteration with sequence-to-sequence neural network

Named Entities are often rare words, and their transliteration across languages has been a challenging task. In this paper, we study a novel technique that segments a named entity into a sequence sub-words or characters. We propose to learn the transliteration mechanism using a sequence-to-sequence neural network. Applying the proposed technique to personal named transliteration on LDC dataset, we show impressive results with more than 10 BLEU score improvement over the competing statistic method on the same corpus.