Researchers have developed Shaer, a new framework for generating classical Arabic poetry that simultaneously controls for semantic meaning, meter subform, and poem length. The system adapts the Yehia-7B model using QLoRA fine-tuning and is trained on an enriched corpus of over 116,000 poems. Shaer demonstrates high accuracy in adhering to poetic constraints, achieving over 95% accuracy for base meter and over 91% for meter subform, significantly outperforming its base model and other evaluated systems. The generated poetry has also been assessed by LLM judges and human evaluators for semantic and literary quality, with no exact copies found in the training data. AI
IMPACT Advances controlled text generation capabilities for specialized linguistic and creative domains.
RANK_REASON This is a research paper detailing a new model and dataset for a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
