ARTIFICIAL INTELLIGENCE AND ENANTIOSEMY: A PILOT EVALUATION OF LARGE LANGUAGE MODEL PERFORMANCE ON CONTRADICTORY WORD MEANINGS IN ENGLISH AND UZBEK

Abstract

This article investigates the extent to which contemporary large language models (LLMs) correctly resolve enantiosemy - the phenomenon whereby a single lexical unit carries two opposite meanings within its semantic structure. Enantiosemy has been described extensively from structural, cognitive, and lexicographic perspectives, yet its treatment by artificial intelligence systems remains largely unexamined, particularly for typologically distant and comparatively lower-resource languages such as Uzbek. Combining a review of the theoretical literature with a small-scale diagnostic experiment, the study administered thirty-two disambiguating test sentences (sixteen English, sixteen Uzbek), each built around a lexical item capable of expressing opposite senses, to a general-purpose LLM (Claude, Anthropic, tested in July 2026). Responses were scored for sense accuracy and analysed qualitatively. The model resolved English enantiosemes with high accuracy (93.75%) when an explicit disambiguating co-text was present, but accuracy fell to 68.75% for Uzbek items, particularly those involving archaic, dialectal, or culturally bound senses, or requiring pragmatic and ironic inference rather than lexical co-text alone. The findings suggest that enantiosemy remains a genuine bottleneck for cross-linguistic natural language understanding and point to concrete directions for improving multilingual sense-disambiguation resources and evaluation benchmarks.