Constrained Semantic Decompression in LLMs through Persian Proverb-Conditioned Story Generation

This article presents a novel approach to proverb-conditioned story generation, framed as a constrained semantic decompression task, using the newly introduced Proverb Aligned Narrative Dataset (PAND), which pairs Persian proverbs with corresponding human-written stories. The study reveals a significant "decompression gap" in current large language models (LLMs), where they demonstrate fluency but struggle to accurately convey the moral and causal structures of proverbs. The findings emphasize the importance of explicit reasoning and iterative refinement in improving narrative generation from abstract concepts, highlighting a critical area for enhancement in LLM applications dealing with cultural and semantic nuances.