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

2026-06-10 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A study on "constrained semantic decompression" in large language models (LLMs) investigates proverb-conditioned story generation, specifically focusing on Persian proverbs. Researchers introduced the Proverb Aligned Narrative Dataset (PAND), which pairs proverbs with human-written stories and explicit meanings. Utilizing a hybrid evaluation framework combining human-calibrated LLM-as-a-Judge with structural metrics, the analysis revealed a persistent "decompression gap." This gap indicates that current LLMs often achieve fluent surface-level text but fail to faithfully instantiate the underlying moral and causal structures of proverbs. The research suggests that explicit reasoning and iterative refinement can partially mitigate these failures, implying that errors stem from translating abstract meaning into narrative form rather than a complete knowledge deficit.

Key takeaway

For NLP engineers developing LLM applications requiring deep cultural understanding or abstract-to-narrative generation, recognize the "decompression gap." Your models may produce fluent text but miss core moral or causal structures. Implement explicit reasoning steps and iterative refinement processes within your prompting strategies to improve semantic fidelity and ensure narratives accurately reflect the source's underlying meaning.

Key insights

LLMs struggle with "constrained semantic decompression," failing to translate abstract cultural knowledge like proverbs into faithful narratives.

Principles

• LLMs exhibit a "decompression gap" in abstract-to-realization tasks.
• Errors often stem from meaning translation, not knowledge absence.

Method

Proverb-conditioned story generation is framed as a constrained semantic decompression task, evaluated via a hybrid LLM-as-a-Judge and structural metrics framework.

In practice

• Employ explicit reasoning to improve abstract meaning instantiation.
• Utilize iterative refinement for mitigating narrative decompression failures.

Topics

• Large Language Models
• Semantic Decompression
• Proverb-Conditioned Generation
• Persian Language
• Narrative Generation
• LLM Evaluation
• Cultural Knowledge

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Moral Semantics Survive Machine Translation: Cross-Lingual Evidence from Moral Foundations Corpora — LLM-based machine translation effectively preserves moral semantics for cross-lingual classification, even with cultural nuances.
• Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach — Multi-aspect data generation and iterative refinement significantly enhance LLM literary translation.
• LLMs for automatic annotation of Mandarin narrative transcripts — LLMs can automate complex discourse-level annotation in non-English speech, significantly reducing time.
• Do Language Models Know Theo Has a Wife? Investigating the Proviso Problem — LLMs align with human judgments on the proviso problem but use shallow pattern matching, not deep pragmatic reasoning.
• Cross-Lingual Empirical Evaluation of Large Language Models for Arabic Medical Tasks — English-centric LLMs exhibit significant performance gaps and reliability issues in Arabic medical tasks.
• AI-assisted cultural heritage dissemination: Comparing NMT and glossary-augmented LLM translation in rock art documents — Glossary-augmented LLMs significantly improve terminology accuracy in specialized translation without sacrificing overall quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.