Machine Transliteration: Leveraging on Third Languages

This paper presents two pivot strategies for statistical machine transliteration, namely system-based pivot strategy and model-based pivot strategy. Given two independent source-pivot and pivot-target name pair corpora, the model-based strategy learns a direct sourcetarget transliteration model while the system-based strategy learns a sourcepivot model and a pivot-target model, respectively. Experimental results on benchmark data show that the systembased pivot strategy is effective in reducing the high resource requirement of training corpus for low-density language pairs while the model-based pivot strategy performs worse than the system-based one.