Integration of an Arabic Transliteration Module into a Statistical Machine Translation System

Proceedings of the Second Workshop on Statistical Machine Translation, pages 17–24, Prague, June 2007. c⃝2007 Association for Computational Linguistics Integration of an Arabic Transliteration Module into a Statistical Machine Translation System Mehdi M. Kashani+, Eric Joanis++, Roland Kuhn++, George Foster++, Fred Popowich+ + School of Computing Science Simon Fraser University 8888 University Drive Burnaby, BC V5A 1S6, Canada mmostafa@sfu.ca popowich@sfu.ca ++ NRC Institute for Information Technology 101 St-Jean-Bosco Street Gatineau, QC K1A 0R6, Canada firstname.lastname@cnrc-nrc.gc.ca Abstract We provide an in-depth analysis of the in- tegration of an Arabic-to-English translit- eration system into a general-purpose phrase-based statistical machine translation system. We study the integration from dif- ferent aspects and evaluate the improve- ment that can be attributed to the integra- tion using the BLEU metric. Our experi- ments show that a transliteration module can help significantly in the situation where the test data is rich with previously unseen named entities. We obtain 70% and 53% of the theoretical maximum improvement we could achieve, as measured by an oracle on development and test sets respectively for OOV words (out of vocabulary source words not appearing in the phrase table). 1 Introduction Transliteration is the practice of transcribing a word or text written in one writing system into an- other writing system. The most frequent candidates for transliteration are person names, locations, or- ganizations and imported words. The lack of a fully comprehensive bilingual dictionary including the entries for all named entities (NEs) renders the task of transliteration necessary for certain natural language processing applications dealing with named entities. Two applications where translitera- tion can be particularly useful are machine transla- tion (MT) and cross lingual information retrieval. While transliteration itself is a relatively well- studied problem, its effect on the aforementioned applications is still under investigation. Transliteration as a self-contained task has its own challenges, but applying it to a real applica- tion introduces new challenges. In this paper we analyze the efficacy of integrating a transliteration module into a real MT system and evaluate the performance. When working on a limited domain, given a suf- ficiently large amount of training data, almost all of the words in the unseen data (in the same do- main) will have appeared in the training corpus. But this argument does not hold for NEs, because no matter how big the training corpus is, there will always be unseen names of people and locations. Current MT systems either leave such unknown names as they are in the final target text or remove them in order to obtain a better evaluation score. None of these methods can give the reader who is not familiar with the source language any informa- tion about those out-of-vocabulary (OOV) words, especially when the source and target languages use different scripts. If these words are not names, one can usually guess what they are, by using the partial information of other parts of speech. But, in the case of names, there is no way to determine the individual or location the sentence …

Content truncated to 500 words