Unsupervised Decomposition of Morphology a Distributed Representation of the Italian Verb System

The paper presents a morphological learning process simulated by an ICA (Independent Component Analysis) algorithm. In an unsupervised manner, the algorithm is able to discover emergent morphologically-motivated features from a representative corpus of Italian verbs. The discovered features can be assumed as non-discrete and distributed representations of the morphological data. Final results also reflect a desirable configuration of an associative network, in which different verbal forms share features or categories corresponding to grammatical and paradigmatic attributes. Without reference to the linguistic notion of ‘morpheme’, our system provides a functional model that explains some peculiar aspects of the organization of the mental lexicon.