The linguistic structure of poetic texts and its formalization for neural machine translation

Abstract

Generative neural networks processing natural language, including large language models designed for neural machine translation, have made remarkable progress, but we are still far from developing artificial intelligence (AI) tools that can translate literary works such as human-produced poetry, resistant to formalization due to its nature, and fully capture their original meaning, style, and imagery. To address these issues, the present study introduces a hybrid evaluation framework that combines traditional methods of linguistic poetics (as developed by Yu.N. Tynyanov, G.O. Vinokur, A.A. Potebnya, M.M. Bakhtin, and others) with those used by modern large language models (DeepSeek-R1 and ChatGPT-5). The framework was applied to three poems by A.S. Pushkin (“I Loved You,” “The Winter Morning,” and “Monument”) to find which formal and semantic features are preserved or lost when poetry is translated from Russian into English by AI. Recommendations for fine-tuning of existing neural network models were made.