Statistical and Syllabification Based Model for Nepali Machine Transliteration

Abstract

Machine Transliteration is one of the important modules for the development of a correct Machine Translation (MT) system. Machine Translation is the technique of converting sentences in one natural language into another using a machine, whereas Machine Transliteration is the method of converting words in one language into phonetically identical words in another. When Machine Translation is unable to translate the Out-of-Vocabulary (OOV) words, Name Entity words, technical words, abbreviation, etc. then Machine Transliteration transliterates these words phonetically. This paper presents a transliteration system for the English-Nepali language pair using the most widely used statistical method with a linguistic syllabification methodology. A model has been designed based on syllable splitting that splits 19,513 parallel entries which contains person names, place, etc. IRSTLM and GIZA++ are used to build the language model (LM) and translation model (TM) i.e. word alignment respectively over parallel entries. For English-Nepali parallel entries on Syllable based split, an accuracy of 87% has been achieved.