Abstract
Multilingual machine translation has recently been in vogue given itspotential for improving machine translation performance for low-resourcelanguages via transfer learning. Empirical examinations demonstrating thesuccess of existing multilingual machine translation strategies, however, arelimited to experiments in specific language groups. In this paper, we present ahierarchical framework for building multilingual machine translation strategiesthat takes advantage of a typological language family tree for enablingtransfer among similar languages while avoiding the negative effects thatresult from incorporating languages that are too different to each other.Exhaustive experimentation on a dataset with 41 languages demonstrates thevalidity of the proposed framework, especially when it comes to improving theperformance of low-resource languages via the use of typologically relatedfamilies for which richer sets of resources are available.