A Purely Monotonic Approach to Machine Translation for Similar Languages

This paper investigates the effect of taking a strictly monotonic approach to machine translation for a restricted set of suitable language pairs. We studied the effect of decoding monotonically for a set of language pairs which has similar word order characteristics and found that for some language pairs - namely language pairs where both languages are in SOV order - there was almost no difference in machine translation quality. The results of this experiment motivated the extension of the monotonic approach into the alignment stage of the training. We used a Bayesian non-parametric aligner that has been shown to out-perform GIZA++ in combination with the grow-diag-final- and heuristic on transliteration data. Our results show that the monotonic aligner was able to match the performance of the GIZA++ baseline, and gains in translation performance were obtained by integrating both aligners into the systems.