This paper presents a structured framework for WordNet synset selection applied to Ancient Greek lexical material. Starting from synonym definitions extracted from the Liddell–Scott–Jones (LSJ) lexicon, we compare two strategies: hierarchy-driven aggregation via bounded hypernym trees and LLM-based definitional matching with pairwise ranking. Graded human evaluation shows that structure-aware methods provide a robust baseline, particularly for nouns and verbs, while LLM-based reranking does not consistently improve performance, especially for highly ploysemous groups of synonyms. Beyond supporting the development of an Ancient Greek WordNet, the study highlights the methodological portability of the framework to other languages and lexical resources.