Name Translation in Statistical Machine Translation - Learning When to Transliterate

Proceedings of ACL-08: HLT, pages 389–397, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Name Translation in Statistical Machine Translation Learning When to Transliterate Ulf Hermjakob and Kevin Knight University of Southern California Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292, USA fulf,knight g@isi.edu Hal Daum´e III University of Utah School of Computing 50 S Central Campus Drive Salt Lake City, UT 84112, USA me@hal3.name Abstract We present a method to transliterate names in the framework of end-to-end statistical machine translation. The system is trained to learn when to transliterate. For Arabic to English MT, we developed and trained a transliterator on a bitext of 7 million sen- tences and Google’s English terabyte ngrams and achieved better name translation accuracy than 3 out of 4 professional translators. The paper also includes a discussion of challenges in name translation evaluation. 1 Introduction State-of-the-art statistical machine translation (SMT) is bad at translating names that are not very common, particularly across languages with differ- ent character sets and sound systems. For example, consider the following automatic translation:1 Arabic input � � A K . � � � � P @ � P � � � p A K . � � J � � �J � J � � �J � � � � � J � � � @ P � � � � � J J � � K A � k P � � � A � � � � � � � � � � � � J J � K . � � �J � J � � � � � � Q K . � SMT output musicians such as Bach Correct translation composers such as Bach, Mozart, Chopin, Beethoven, Schumann, Rachmaninoff, Ravel and Prokofiev The SMT system drops most names in this ex- ample. “Name dropping” and mis-translation hap- pens when the system encounters an unknown word, mistakes a name for a common noun, or trains on noisy parallel data. The state-of-the-art is poor for 1taken from NIST02-05 corpora two reasons. First, although names are important to human readers, automatic MT scoring metrics (such as BLEU) do not encourage researchers to improve name translation in the context of MT. Names are vastly outnumbered by prepositions, articles, adjec- tives, common nouns, etc. Second, name translation is a hard problem — even professional human trans- lators have trouble with names. Here are four refer- ence translations taken from the same corpus, with mistakes underlined: Ref1 composers such as Bach, missing name Chopin, Beethoven, Shumann, Rakmaninov, Ravel and Prokoviev Ref2 musicians such as Bach, Mozart, Chopin, Bethoven, Shuman, Rachmaninoff, Rafael and Brokoviev Ref3 composers including Bach, Mozart, Schopen, Beethoven, missing name Raphael, Rahmaniev and Brokofien Ref4 composers such as Bach, Mozart, missing name Beethoven, Schumann, Rachmaninov, Raphael and Prokofiev The task of transliterating names (independent of end-to-end MT) has received a significant amount of research, e.g., (Knight and Graehl, 1997; Chen et al., 1998; Al-Onaizan, 2002). One approach is to “sound out” words and …

Content truncated to 500 words