ISM@FIRE-2013 Shared Task on Transliterated Search

This paper describes the approach we adopted during official submission of FIRE-2013 Shared Task on Transliterated Search along with few other approaches that we experimented post-submission. The techniques solve the problem of language labeling, by identifying query word as English or Hindi (E or H) term in mixed language sentence queries. Manual and machine learning algorithms are used. For the transliteration of H labeled word we use manual (dictionary based), generative (grapheme based) and combination of both in different algorithms. We observe that learning based classification improves labeling accuracy. Extraction based transliteration gives better result than Generation based when the terms are available in bilingual dictionary. But it may lead to incorrect transliteration if terms are wrongly aligned and the approach fails for out-of-dictionary words. In this case transliteration by generation is the only alternative. But generation alone does not perform well because of spelling variation in transliterated terms. During evaluation we also observe that transliteration systems are generally corpus-biased. Although our performance in the official submission was moderate, we obtain better results during our post-submission experiments.