Treating words as opaque symbols is the wrong assumption for morphologically complex languages. Encoding morphology into three model types — a Bayesian compound-decomposition n-gram model, a distributed model learning vector representations of morphemes, and an unsupervised model that discovers even discontiguous (Semitic-style) morphemes — improves model performance and beats the data sparsity that morphological processes create, with benefits for machine translation.

— In our words · Source: Probabilistic Modelling of Morphologically Rich Languages (Jan A. Botha)