Abstract
We study unsupervised multilingual alignment, the problem of findingword-to-word translations between multiple languages without using any paralleldata. One popular strategy is to reduce multilingual alignment to the muchsimplified bilingual setting, by picking one of the input languages as thepivot language that we transit through. However, it is well-known thattransiting through a poorly chosen pivot language (such as English) mayseverely degrade the translation quality, since the assumed transitiverelations among all pairs of languages may not be enforced in the trainingprocess. Instead of going through a rather arbitrarily chosen pivot language,we propose to use the Wasserstein barycenter as a more informative "mean"language: it encapsulates information from all languages and minimizes allpairwise transportation costs. We evaluate our method on standard benchmarksand demonstrate state-of-the-art performances.