Statistical Machine Transliteration with Multi-to-Multi Joint Source Channel Model

This paper describes DFKI’s participation in the NEWS2011 shared task on machine transliteration. Our primary system participated in the evaluation for English-Chinese and Chinese-English language pairs. We extended the joint sourcechannel model on the transliteration task into a multi-to-multi joint source-channel model, which allows alignments between substrings of arbitrary lengths in both source and target strings. When the model is integrated into a modified phrasebased statistical machine translation system, around 20% of improvement is observed. The primary system achieved 0.320 on English-Chinese and 0.133 on Chinese-English in terms of top-1 accuracy.