Urdu Hindi Machine Transliteration using SMT

Transliteration is a process of transcribing a word of the source language into the target language such that when the native speaker of the target language pronounces it, it sounds as the native pronunciation of the source word. Statistical techniques have brought significant advances and have made real progress in various fields of Natural Language Processing (NLP). In this paper, we have analysed the application of Statistical Machine Translation (SMT) for solving the problem of Urdu Hindi transliteration using a parallel lexicon. We have designed total 24 Statistical Transliteration (ST) systems by combining different types of alignments, translation models and target language models. We have performed total 576 experiments and have reported significant results. From Hindi–to–Urdu transliteration, we have achieved the maximum word-level accuracy of 71.5%. From Urdu–to–Hindi transliteration, the maximum word-level accuracy is 77.8% when the input Urdu text contains all necessary diacritical marks and 77% when the input Urdu text does not contain all necessary diacritical marks. At character-level, transliteration accuracy is more than 90% in both directions.