This study explores the extent to which large-language models (LLMs) can effectively translate poetic works from Italian to English while maintaining alliterative fidelity and semantic accuracy. We chose to translate poems from Eugenio Montale, 1975 Nobel Prize in Literature laureate, who is known for the prominent use of sound devices and alliteration in his poetry collection Ossi di seppia [Cuttlefish bones]. To supplement the translation’s creation through the retrieval-augmented generation (RAG) technique, the LLM (Meta Llama 3 8B Instruct) is given context from literary analyses and other resources pertaining to Montale and his poems.

To observe the effect of providing incremental context on translation quality, the RAG system is queried three times upon adding a new document to the system’s vector database to ensure that the LLM outputs translations of consistent quality. We then analyze each translation to determine areas of improvement and use them to refine the system instructions through prompt engineering until the LLM consistently provides a satisfactory translation. The resulting prompt, which may be supplemented with additional documents from the vector database, will then be used as a baseline to derive translations from Montale’s other poems to verify whether they can output translations of acceptable quality.