Disambiguated linear word translation in medium European languages

Abstract

An earlier paper used triangulated word translations as seed in linear translation between medium European languages. The present work improves upon it by handling word ambiguity both in the main (i.e. source and target) languages and in the pivot by training a multi-prototype vector space model in the former, filtering triangles based on scores computed by a linear model trained with direct (non-triangulated) translations, and finally translating the whole vocabulary with a linear translation trained on filtered triangles (in addition to direct ones).