Semantic non-compositional exploration of Bessarabian idioms by LLMs

Abstract

The study explores the potential of LLMs in interpreting and translating Bessarabian idioms. The central problem addressed is the semantic non-compositionality of idiomatic expressions, which poses a significant challenge for Natural Language Processing since their figurative meaning cannot be derived from literal components. As part of the CI ARiA project, 1000 proverbs were digitized, using a corpus of 400 of them to evaluate the performance of 10 AI models (such as ChatGPT, Gemini, Grok). The methodology is multi-algorithmic, combining textual distance metrics (Levenshtein, Jaccard) with semantic similarity analysis via Sentence Transformers. The results indicate that while models demonstrate a solid capacity to grasp metaphorical meanings, significant differences exist regarding consistency and explanatory style.