English to Hindi Machine Transliteration System at NEWS 2009

Abstract

This paper reports about our work in the NEWS 2009 Machine Transliteration Shared Task held as part of ACL-IJCNLP 2009. We submitted one standard run and two non-standard runs for English to Hindi transliteration. The modified joint source-channel model has been used along with a number of alternatives. The system has been trained on the NEWS 2009 Machine Transliteration Shared Task datasets. For standard run, the system demonstrated an accuracy of 0.471 and the mean F-Score of 0.861. The non-standard runs yielded the accuracy and mean F-scores of 0.389 and 0.831 respectively in the first one and 0.384 and 0.828 respectively in the second one. The non-standard runs resulted in substantially worse performance than the standard run. The reasons for this are the ranking algorithm used for the output and the types of tokens present in the test set.