Transliteration Retrieval Model for Cross Lingual Information Retrieval

The performance of transliteration from a source language to a target language builds the ground work in support of proper name Cross Lingual Information Retrieval (CLIR). Traditionally, this task is accomplished by two separate modules: transliteration and retrieval. Queries are first transliterated to target language using one or multiple hypotheses. The retrieval is then carried out based on translated queries. The transliteration often results in 30-50% errors with top I hypothesis, thus leading to significant performance degradation in CUR. Therefore, we proposed a unified transliteration retrieval model that incorporates the transliteration similarity measurement into the relevance scoring function. In addition, we presented an efficient and robust method in similarity measurement for a given proper name pair using the Hidden Markov Model (HMM) based alignment and a Statistical Machine Translation (SMT) framework. Experimental data showed significant results with the proposed integrated method on the NTCIR7 IR4QA task, which demonstrated a greater flexibility and acceptance in transliteration.