English to Bangla Phrase-Based Machine Translation

English to Bangla Phrase-Based Machine Translation Md. Zahurul Islam Dipartimento di Informatica Universit`a di Pisa islam@di.unipi.it J¨org Tiedemann Department of Linguistics and Philology Uppsala University jorg.tiedemann@lingfil.uu.se Andreas Eisele Language Technology Lab DFKI GmbH, Saarbr¨ucken eisele@dfki.de Abstract Machine Translation (MT) is the task of automatically translating a text from one language to another. In this work we de- scribe a phrase-based Statistical Machine Translation (SMT) system that translates English sentences to Bangla. A translit- eration module is added to handle out- of-vocabulary (OOV) words. This is es- pecially useful for low-density languages like Bangla for which only a limited amount of training data is available. Fur- thermore, a special component for han- dling preposition is implemented to treat systematic grammatical differences be- tween English and Bangla. We have shown the improvement of our system through effective impacts on the BLEU, NIST and TER scores. The overall BLEU score of our system is 11.7 and for short sentences it is 23.3. 1 Introduction SMT requires enormous amount of parallel text in the source and target language to achieve high quality translation. However, many languages are considered to be low-density languages, either be- cause the population speaking the language is not very large, or because insufficient digitized text material is available in a language even though it is spoken by millions of people. Bangla/Bengali is one such language. Bangla, an Indo-Aryan lan- guage, is a language of Southeast Asia, which comprises present day Bangladesh and the Indian state of West Bengal. With nearly 230 million speakers, Bangla is one of the most spoken lan- guages in the world, but only a very small number of tools and resources are available for Bangla. Our aim in this work is to present a phrase- based SMT system for translating English sen- tences to Bangla. The current state-of-the-art phrase-based SMT systems available for this task is based on a log-linear translation model, which is used as our baseline system. Though we have built our system with a small amount of training data compared to MT systems of other language pairs, we have got better results than existing MT systems for this language pair when tested on the same domain. A transliteration module has been added as a component with the translation system to handle OOV words. For the transliteration mod- ule we applied the same phrase-based SMT model, but using characters instead of words. Another difference to standard SMT models is the prepo- sition handling module we added to our system to handle prepositional diversity between English and Bangla. Instead of prepositions, as in English, Bangla uses postpositions, or attaches inflections to the head noun of the prepositional phrase (i.e., the object of the preposition) (Naskar and Bandy- opadhyay, 2006b). The rest of the paper is organized as follows. Section 2 summarizes some related past work. Section 3 shows the importance of preposition handing during translation. Section 4 describes the experiments and results. Finally, we summarize our observation, and outline future work in Sec- tion 5, …

Content truncated to 500 words