Pivot translation using source-side dictionary and target-side parallel corpus towards MT from resource-limited languages

Statistical machine translation (SMT) requires a parallel corpus between the source and target languages. This requirement makes SMT difficult to apply to resource-limited languages that do not have any parallel corpora even to a major language, e.g., English. For such a problem, a novel pivot translation method has been proposed that does not require the source-side parallel corpus, but, uses a word dictionary instead. In this paper, we evaluate the relative translation performance of the dictionary-based method by comparing it with both the standard SMT that uses a direct parallel corpus, and the conventional pivot translation that uses two parallel corpora, by using the Europarl corpus. In addition, we also investigate the edge weighting and lattice pruning methods applied to the word lattice that was used to represent the pivot sentence candidates in the dictionary-based method.