Statistical Machine Translation (SMT) typically has difficulties with less-resourced languages even with homogeneous data. In this thesis we address the application of Example-Based Machine Translation (EBMT) methods to overcome some of these difficulties. We adopt three alternative approaches to tackle these problems focusing
on two poorly-resourced translation tasks (English–Bangla and English–Turkish). First, we adopt a runtime approach to EBMT using proportional analogy. In addition to the translation task, we have tested the EBMT system using proportional analogy for named entity transliteration. In the second attempt, we use a compiled approach to EBMT. Finally, we present a novel way of integrating Translation Memory (TM) into an EBMT system. We discuss the development of these three different EBMT systems and the experiments we have performed. In addition, we present an approach to augment the output quality by strategically combining EBMT systems and SMT systems. The hybrid system shows significant improvement for different language pairs.
Runtime EBMT systems in general have significant time complexity issues especially for large example-base. We explore two methods to address this issue in our system by making the system scalable at runtime for a large example-base (English–French). First, we use a heuristic-based approach. Secondly we use an IR-based indexing technique to speed up the time-consuming matching procedure of the EBMT system. The index-based matching procedure substantially improves run-time speed without affecting translation quality.