Acquisition and Representation of Grammatical Categories: Grammatical Gender in a Connectionist Network

In traditional models of language production grammatical categories are represented as abstract features independent of semantics and phonology. An alternative view is proposed where syntactic categories emerge as a higher-order regularity from semantic and phonological properties of words. The proposal was tested using grammatical gender in Serbian, a south Slavic language with rich morphology. Semantic and phonological correlates of gender are described using a corpus of 1221 Serbian nouns. A PDP network was trained to produce the same words based on distributed semantic representation as input and distributed phonological representation as output, and with no explicit representation of grammatical gender. Upon successful learning of the training corpus, generalization was explored using test corpora designed to capture semantic and phonological properties of different genders. The findings suggest that grammatical gender, as other syntactic categories, may be viewed as emerging through coherent co-variation of semantic and phonological properties of words during learning.