Morphosyntactic Target Language Matching in Statistical Machine Translation

Morphosyntactic Target Language Matching in Statistical Machine Translation Simon Zwarts Centre for Language Technology Macquarie University Sydney, Australia szwarts@ics.mq.edu.au Mark Dras Centre for Language Technology Macquarie University Sydney, Australia madras@ics.mq.edu.au Abstract While the intuition that morphological pre- processing of languages in various applica- tions can be beneficial appears to be often true, especially in the case of morphologi- cally richer languages, it is not always the case. Previous work on translation between Nordic languages, including the morphologi- cally rich Finnish, found that morphological analysis and preprocessing actually led to a decrease in translation quality below that of the unprocessed baseline. In this paper we investigate the proposition that the effect on translation quality depends on the kind of morphological preprocessing; and in particular that a specific kind of mor- phological preprocessing before translation could improve translation quality, a prepro- cessing that first transforms the source lan- guage to look more like the target, adapted from work on preprocessing via syntactically motivated reordering. We show that this is in- deed the case in translating from Finnish, and that the results hold for different target lan- guages and different morphological analysers. 1 Introduction In many NLP applications, morphological prepro- cessing such as stemming is intuitively felt to be im- portant, especially when dealing with morphologi- cally rich languages. In fairly early work on mor- phological disambiguation of the agglutinative lan- guage Turkish, Hakkani-Tur et al. (2000) cite the note of Hankamer (1989) that a single Turkish root can potentially have over a million inflected vari- ants; such a proliferation of forms could exacer- bate any data sparsity problems. Zwarts and Dras (2007a) showed, for languages that differ in mor- phological richness but are otherwise structurally similar, that, as expected, morphological preprocess- ing in an Statistical Machine Translation (SMT) en- vironment confers greater benefits on the morpho- logically richer language than on the morphologi- cally poorer one. However, it is not always the case that morpholog- ical preprocessing of a morphologically rich lan- guage does provide a benefit. Again in an SMT context, Virpioja et al. (2007) preprocessed Finnish (a Uralic language, typologically between inflected and agglutinative, and consequently morphologi- cally rich) for translation into Danish and Swedish (both Indo-European languages, quite different in many morphosyntactic respects from Finnish, in- cluding being morphologically less rich). In doing this, they found that translation quality was gener- ally worse than baseline, and never better. What this paper looks at is whether it is the way in which the morphology is preprocessed that is important. We draw on an idea from other work in machine translation first presented by Collins et al. (2005), where the source language is reordered so that its syntax is more like that of the target language, leading to an improvement in translation quality; here, we see whether that idea can apply at the level of morphology. In particular, where there are phenomena that are handled via morphology in one language and by syntax in another, we investi- gate whether an adaptation …

Content truncated to 500 words