Guzman-Olivares Daniel Quijano-Sanchez Lara Liberatore Federico

Abstract

The rise of generative chat-based Large Language Models (LLMs) over the pasttwo years has spurred a race to develop systems that promise near-humanconversational and reasoning experiences. However, recent studies indicate thatthe language understanding offered by these models remains limited and far fromhuman-like performance, particularly in grasping the contextual meanings ofwords, an essential aspect of reasoning. In this paper, we present a simple yetcomputationally efficient framework for multilingual Word Sense Disambiguation(WSD). Our approach reframes the WSD task as a cluster discrimination analysisover a semantic network refined from BabelNet using group algebra. We validateour methodology across multiple WSD benchmarks, achieving a new state of theart for all languages and tasks, as well as in individual assessments by partof speech. Notably, our model significantly surpasses the performance ofcurrent alternatives, even in low-resource languages, while reducing theparameter count by 72%.