Improving Low-Resource Statistical Machine Translation with a Novel Semantic Word Clustering Algorithm

In this paper we present a non-languagespecific strategy that uses large amounts of monolingual data to improve statistical machine translation (SMT) when only a small parallel training corpus is available. This strategy uses word classes derived from monolingual text data to improve the word alignment quality, which generally deteriorates significantly because of insufficient training. We present a novel semantic word clustering algorithm to generate the word classes motivated by the word similarity metric presented in (Lin, 1998). Our clustering results showed this novel word clustering outperforms a state-of-the-art hierarchical clustering. We then designed a new procedure for using the derived word classes to improve word alignment quality. Our experiments showed that the use of the word classes can recover over 90% of the loss resulting from the alignment quality that is lost due to the limited parallel training.