Unsupervised Morphological Analysis by Formal Analogy

While classical approaches to unsupervised morphology acquisition often rely on metrics based on information theory for identifying morphemes, we describe a novel approach relying on the notion of formal analogy. A formal analogy is a relation between four forms, such as: reader is to doer as reading is to doing. Our assumption is that formal analogies identify pairs of morphologically related words. We first describe an approach which simply identifies all the formal analogies involving words in a lexicon. Despite its promising results, this approach is computationally too expensive. Therefore, we designed a more practical system which learns morphological structures using only a (small) subset of all formal analogies. We tested those two approaches on the five languages used in Morpho Challenge2009.