Using Example-Based MT to Support Statistical MT when Translating Homogeneous Data in a Resource-Poor Setting

Using Example-Based MT to Support Statistical MT when Translating Homogeneous Data in a Resource-Poor Setting Sandipan Dandapat1, Sara Morrissey1, Andy Way1, Mikel L. Forcada2 1CNGL, School of Computing Dublin City University, Glasnevin, Dublin 9, Ireland {sdandapat, smorri, away}computing.dcu.ie 2Departament de Llenguatges i Sistemes Informàtics Universitat d'Alacant, E-03071 Alacant, Spain mlf@dlsi.ua.es Abstract In this paper, we address the issue of ap- plying example-based machine transla- tion (EBMT) methods to overcome some of the difficulties encountered with stat- istical machine translation (SMT) tech- niques. We adopt two different EBMT approaches and present an approach to augment output quality by strategically combining both EBMT approaches with the SMT system to handle issues arising from the use of SMT. We use these ap- proaches for English to Turkish transla- tion using the IWSLT09 dataset. Im- proved evaluation scores (4% relative BLEU improvement) were achieved when EBMT was used to translate sen- tences for which SMT failed to produce an adequate translation. 1 Introduction Over the past two decades, machine translation (MT) has shown very promising results, mostly using statistical machine translation (SMT) tech- niques. However, a large number of languages exist which suffer from the scarcity of parallel corpora, e.g. Indic languages, sign languages etc. SMT approaches have yielded low translation quality for these poorly resourced languages (Khalilov et al., 2010). It is often the case that domain-specific translation is required to tackle the issue of scarce resources, but it can still suf- fer from very low accuracy within the SMT framework, even for homogeneous domains (Dandapat et al., 2010). Although SMT and EBMT are both data-driven approaches to MT, both of them have their own advantages and lim- © 2011 European Association for Machine Translation. itations. Typically, an SMT system works well with significant amounts of training data. In con- trast, an EBMT approach can be developed with a limited example-base (Somers, 2003); also, as with any other data-driven system, an EBMT system works well when training and test sets are quite close in nature (Marcu, 2001). This is be- cause EBMT systems reuse the segments of test sentences that can be found in the source side of the example-base at runtime. Keeping these points in mind, it is important to develop an MT system of reasonably good quality based on limited amounts of data. In this direction, we are inspired to examine different EBMT approaches which can handle the problem of data sparseness. It is often the case that EBMT systems produce a good translation where SMT fails and vice versa. In order to harness the ad- vantages of both approaches, we use a careful combination of both EBMT and SMT to improve translation accuracy. We adopt two alternative approaches to tackle the above problems. First we adopt a compiled approach to EBMT which essentially produces translation templates during the training stage, based on the description in (Cicekli and Güvenir, 2001). Our second attempt presents a novel way of integrating translation memory (TM) into an EBMT system. Starting with the user’s TM …

Content truncated to 500 words