A study of word-classing for MT reordering

Abstract

MT systems typically use parsers to help reorder constituents. However most languages do not have adequate treebank data to learn good parsers, and such training data is extremely time-consuming to annotate. Our earlier work has shown that a reordering model learnt from word-alignments using POS tags as features can improve MT performance (Visweswariah et al., 2011). In this paper, we investigate the effect of word-classing on reordering performance using this model. We show that unsupervised word clusters perform somewhat worse but still reasonably well, compared to a part-of-speech (POS) tagger built with a small amount of annotated data; while a richer tagset including case and gender-number-person further improves reordering performance by around 1.2 monolingual BLEU points. While annotating this richer tagset is more complicated than annotating the base tagset, it is much easier than annotating treebank data.