Transition-Based Morphological Disambiguation

In Morphologically Rich Languages (MRLs), sentences are composed of ambiguous space-delimited tokens that ought to be disambiguated with respect to their constituent morphemes. Previous work on Morphological Disambiguation (MD) of MRLs has had variable success, with Semitic languages having sub-par results for downstream applications. Here we propose novel MD transition-based systems, both word-based and morphemebased, and tackle the challenge introduced by the variable length of hypothesized morpheme sequences. Our experiments show that transition-based morphemebased MD consistently outperforms the word-based variant, while providing new state of the art results on Hebrew MD.