Language Independent Named Entity Transliteration

摘要

This paper reports about a modified joint source -channel model that has been used along with a number of alternatives to generate English transliterations of the Bengali, Hindi and Telugu Named entities (NE). The Bengali, Hindi and Telugu NEs are divided into Transliteration Units (TUs) that have the pattern C{sup}+M, where C represents a vowel or a consonant or conjunct and M represents the vowel modifier or matra. An English word is divided into TUs that have the pattern C{sup}*V{sup}*, where C represents a consonant and V represents a vowel. The system learns mappings automatically from the bilingual training sets of NEs, particularly person and location names. The output of this mapping process is a decision-list classifier with collocated TUs in the source language and their equivalent TUs in collocation in the target language along with the probability of each decision obtained from the training set. Evaluation of the proposed models in terms of Word Agreement Ratio (WAR) and Transliteration Unit Agreement Ratio (TUAR) demonstrated that the modified joint source-channel model performs best for Bengali, Hindi and Telugu to English transliterations. All the transliteration models have been tested for English to Bengali transliteration and it has been observed that the modified joint source-channel performs best in terms of evaluation parameters, WAR and TUAR. The Bengali to English (B2E) and English to Bengali (E2B) transliteration systems have also been tested with the help of linguistic knowledge in the form of conjuncts and diphthongs in Bengali and their corresponding representations in English. It has been observed that with the inclusion of linguistic knowledge the modified joint source-channel model performs best.