Linguistic Efficiency and Inhibitory Control: Evaluating the DIINA Model on Afrikaans as a Morphologically Simplified but Context-Rich Language

Abstract Afrikaans presents a compelling test case for computational models of negation and ambiguity resolution: it is morphologically relatively streamlined, yet it relies heavily on context, syntax, and discourse-level interpretation to disambiguate meaning. This article evaluates DIINA, an inhibitory neural architecture designed for double negation resolution and lexical ambiguity reduction in Afrikaans. DIINA operationalizes inhibitory control as a learned mechanism that suppresses competing interpretations when negation is structurally and semantically licensed. The central claim is that linguistic efficiency in Afrikaans does not arise from morphological richness alone, but from a tight coupling between local cues and broader contextual integration. Empirically, DIINA achieved 94% accuracy on double negation and reduced lexical ambiguity error to 6% (pasted-text.txt:1311– 1316; 1529; 1736). In the sentence “Ek sien nie die man nie”, the secondary negation token received an inhibition score of 0.92, and inhibitory effects intensified in the final layers of the network, indicating progressive suppression of unlicensed readings (pasted-text.txt:1241). To support interpretability, the model uses heatmaps to visualize inhibition patterns and layer wise attribution (pasted-text.txt:812–814; 835–836). The findings suggest that Afrikaans is particularly well suited for studying how neural architectures can learn to resolve semantic conflicts through inhibitory dynamics rather than only through surface form morphology. The article situates DIINA within broader work on negation, ambiguity, and neural interpretability, and argues that inhibitory modeling offers a promising route toward cognitively plausible and linguistically informed NLP for context-rich languages.