Marathi - English Named Entities Forward MachineTransliteration Using Linguistic and Metrical Approach

Machine transliteration is required in machine translation to transliterate words which are not present in dictionary such as names of persons, locations, cities and villages, names of roads and building etc. It is noticed that the machine transliteration is less studied for the language pair Marathi-English. Most of the present research in this area is done by using linguistic rule based and grapheme/phoneme based statistical approaches with the help of machine learning tools. This paper proposes Marathi to English forward machine transliteration of named entities of Indian-origin using hybrid approach which is the combination of linguistic rules and metrical approach. Proposed rule-based transliteration method uses phoneme model of the machine transliteration which uses the phonetic mapping between the source language Marathi written using Devanagari script and target language English written using Roman script. The key concept of this method is schwa deletion based on the stress of syllable after phonetic mapping. This approach requires full consonant based input and uses syllable based metrical approach for schwa deletion. Metrical approach works on stress of syllable whether it is unstressed or stressed in the given named entity. The experimental result showed that the absolute performance of the proposed method is high as compare to existing grapheme based approaches.