- Main
The Integrated Information Bottleneck: A theoretically-grounded model unifying category- and system-level structure in communicative efficiency
Abstract
Much research has shown that lexical semantic systems are shaped by a drive for efficient communication, which yields desirable structure at the level of individual lexical categories (i.e., the semantic coherence of words). Less attention has been given to what we call semantic systematicity: structural regularity that exists across lexical categories, where different words make the same semantic distinctions. We develop the first model of the efficiency trade-off that captures both category- and system-level regularity using the same domain-general information-theoretic underpinning. We also propose a stricter evaluation method for comparing alternative formalizations of the efficiency trade-off. Our model that integrates the pressures for category-level structure and semantic systematicity explains crosslinguistic variation better than previous approaches in two semantic domains, shedding new light on how a pressure for efficiency shapes the structure of language.