Abstract
We consider the problem of aligning continuous word representations, learnedin multiple languages, to a common space. It was recently shown that, in thecase of two languages, it is possible to learn such a mapping withoutsupervision. This paper extends this line of work to the problem of aligningmultiple languages to a common space. A solution is to independently map alllanguages to a pivot language. Unfortunately, this degrades the quality ofindirect word translation. We thus propose a novel formulation that ensurescomposable mappings, leading to better alignments. We evaluate our method byjointly aligning word vectors in eleven languages, showing consistentimprovement with indirect mappings while maintaining competitive performance ondirect word translation.