A Cascaded Framework for Statistical Machine Translation System Combination

This paper investigates an extensive evaluation of combination techniques and presents a cascaded framework for combining multiple machine translation (MT) system outputs. A word transition network (WTN) is constructed from an N -best list by aligning the hypotheses against an alignment reference, where the alignment is based on minimising an modified translation edit rate (TER) with word or phrase reordering. The minimum Bayes risk (MBR) decoding techinque is inverstigated for the selection of an appropriate alignment reference. Pairwise word alignment is created by an enhanced statistical alignment algorithm that explicitly models word reordering. Experimental results are presented based on three MT systems of Chinese-English translation outputs. It is shown that worthwhile improvements in translation performance can be obtained using the proposed framework.