GanjNet: Leveraging Network Modeling with Large Language Models for Persian Word Sense Induction

This paper introduces GanjNet, a novel approach to Word Sense Induction (WSI) in the Persian language that leverages network modeling and community detection in conjunction with large language models (LLMs). We present a method that constructs semantic graphs from lexical substitutes generated by LLMs and applies community detection algorithms to uncover and distinguish word senses in unannotated text. GanjNet addresses challenges such as limited annotated resources, high degrees of polysemy, and context-sensitive meanings in Persian. By leveraging unsupervised techniques, we enhance sense induction without relying on extensive labeled data. Our experiments demonstrate that GanjNet outperforms existing methods on a custom dataset derived from MirasText, achieving a V-measure of 47% and a paired F-score of 58%, compared to the best baseline method with a V-measure of 41% and a paired F-score of 53%. These results showcase the potential of integrating community detection and LLMs for unsupervised semantic tasks in morphologically rich languages like Persian. Moreover, GanjNet's flexibility offers practical applicability across various domains, including automatic thesaurus and WordNet generation, as well as assisting writers in context-sensitive word choice, demonstrating its broader impact on natural language understanding.