Most recent approaches to bilingual dictionaryinduction find a linear alignment between theword vector spaces of two languages. Weshow that projecting the two languages ontoa third, latent space, rather than directly ontoeach other, while equivalent in terms of expressivity,makes it easier to learn approximatealignments. Our modified approach also allowsfor supporting languages to be includedin the alignment process, to obtain an even betterperformance in low resource settings.