Neural Machine Translation for Poetry: Balancing Semantic Fidelity and Rhythmic Aesthetics Using Transformer Models

We propose a Transformer-based model that includes a semantic alignment module to ensure accurate meaning retention and multi-head rhythm attention mechanism to capture rhythmic components such as syllable stress, rhyme, and meter. Additionally, a dual-objective optimization technique is developed throughout the translation process to dynamically reconcile the opposing requirements of rhythmic and semantic constraints. Experimental results from a bilingual poetry dataset show that the proposed model considerably outperforms state-of-the-art baselines in terms of BLEU scores, semantic similarity, and rhythmic consistency, as compared to mBART and classic Transformer designs. This method ensures that the semantic integrity of the original text is preserved while producing high-quality translations that accurately portray its aesthetic parts and structural aspects.