Name transliteration with bidirectional perceptron edit models

Abstract

We report on our efforts as part of the shared task on the NEWS 2009 Machine Transliteration Shared Task. We applied an orthographic perceptron character edit model that we have used previously for name transliteration, enhancing it in two ways: by ranking possible transliterations according to the sum of their scores according to two models, one trained to generate left-to-right, and one right-to-left; and by constraining generated strings to be consistent with character bigrams observed in the respective language's training data. Our poor showing in the official evaluation was due to a bug in the script used to produce competition-compliant output. Subsequent evaluation shows that our approach yielded comparatively strong performance on all alphabetic language pairs we attempted.