Sinhala-Tamil Machine Translation: Towards better Translation Quality

Sinhala-Tamil Machine Translation: Towards better Translation Quality Randil Pushpananda Ruvan Weerasinghe Language Technology Research Laboratory University of Colombo School of Computing Sri Lanka {rpn|arw}@ucsc.lk Mahesan Niranjan School of Electronics and Computer Science University of Southampton, UK mn@ecs.soton.ac.uk Abstract Statistical Machine Translation (SMT) is a well-known and well established data- driven approach used for language trans- lation. The focus of this work is to de- velop a statistical machine translation sys- tem for Sri Lankan languages, Sinhala and Tamil language pair. This paper presents a systematic investigation of how Sinhala- Tamil SMT performance varies with the amount of parallel training data used, in order to find out the minimum needed to develop a machine translation system with acceptable performance. 1 Introduction Sri Lanka is a multi-ethnic, multi-lingual country. Sinhala and Tamil are the national languages of Sri Lanka. The majority of Sri Lankans do not have a good knowledge of languages other than their mother tongue. Therefore a language barrier between the Sinhala and Tamil communities ex- ists. This language barrier and the problems that arose during the last 30 years in the country, en- couraged us to a translation application using the SMT approach. This would reduce the language gap between these two communities and thereby help solve a burning issue in the country. The choice of the Sinhala - Tamil language pair provides some opportunities as well as some challenges. The opportunity is that they share some affinity to each other, having evolved along- side each other in Sri Lanka. The challenges in- clude the sparseness in the availability of data, and the limited research undertaken in them. Hence, developing a successful system with limited re- sources is our ultimate goal. 2 Background and Related Work There is very limited research reported in the lit- erature for Sinhala-Tamil machine translation. Ac- cording to (Weerasinghe, 2003), the Sinhala-Tamil language pair gives better performance compared to the Sinhala-English pair in SMT since they are more closely related to each other owing to their evolution within Sri Lanka. Some impor- tant factors to consider when building SMT for the Sinhala-Tamil language pair have been identified in (Sakthithasan et al., 2010). The limited amount of data, and the restricted domain it represented, makes that word hard to generalize. Another study (Jeyakaran and Weerasinghe, 2011), explored the applicability of the Kernel Ridge Regression tech- nique to Sinhala-Tamil translation. This research resulted in a hybrid of classical phrase based SMT and Kernel Ridge Regression with two novel solu- tions for the pre-image problem. Owing to the limited amount of parallel data available, it has been not possible to analyze how the results vary with increasing numbers of paral- lel sentences in Sinhala and Tamil for general pur- pose MT. 2.1 Sinhala and Tamil Languages Sinhala belongs to the Indo-Aryan language fam- ily and Tamil to the Dravidian family. Both Sin- hala and Tamil languages are morphologically rich languages: Sinhala has up to 110 noun word forms and up to 282 verb word forms (Welgama et …

Content truncated to 500 words