Statistically Motivated Example-based Machine Translation using Translation Memory

In this paper we present a novel way of integrating Translation Memory into an Example-based Machine Translation System (EBMT) to deal with the issue of low resources. We have used a dialogue of 380 sentences as the example-base for our system. The translation units in the Translation Memories are automatically extracted based on the aligned phrases (words) of a statistical machine translation (SMT) system. We attempt to use the approach to improve translation from English to Bangla as many statistical machine translation systems have difficulty with such small amounts of training data. We have found the approach shows improvement over a baseline SMT system.