Don't Go Far Off: An Empirical Study on Neural Poetry Translation

Abstract

Despite constant improvements in machine translation quality, automatic poetry translation remains a challenging problem due to the lack of open-sourced parallel poetic corpora, and to the intrinsic complexities involved in preserving the semantics, style and figurative nature of poetry. We present an empirical investigation for poetry translation along several dimensions: 1) size and style of training data (poetic vs. non-poetic), including a zeroshot setup; 2) bilingual vs. multilingual learning; and 3) language-family-specific models vs. mixed-language-family models. To accomplish this, we contribute a parallel dataset of poetry translations for several language pairs. Our results show that multilingual fine-tuning on poetic text significantly outperforms multilingual fine-tuning on non-poetic text that is 35X larger in size, both in terms of automatic metrics (BLEU, BERTScore, COMET) and human evaluation metrics such as faithfulness (meaning and poetic style). Moreover, multilingual fine-tuning on poetic data outperforms bilingual fine-tuning on poetic data. 1

Rhythmic lyrics translation: Customizing a pre-trained language model using stacked fine-tuning

“…We could not find any studies on translating lyrics, let alone doing so while maintaining rhythmic characteristics. However, we could find past studies on poetry translation, which is like lyrics in the sense that both have rhythm, rhymes, and sensory details that cannot be inferred by conventional NMT models (3)(4)(5). Researchers first used rhythm and rhyme constrained decoding methods, but the translations often failed to keep the semantics of the original poetry (3,4).…”

“…Researchers first used rhythm and rhyme constrained decoding methods, but the translations often failed to keep the semantics of the original poetry (3,4). To preserve the semantics and figurative characteristics, other researchers used a pre-trained language model and fine-tuned the model with poetic data on multiple language pairs (5).…”

“…For training NMT models, we needed a certain type of dataset known as parallel data, which are in the form of multiple sets of sentences and their translations (6). As neural machine translation has not yet been globally explored, it was incredibly difficult to find parallel data for less popular languages such as Korean, especially for specific domain datasets-in our case-lyrics (5,6). The second problem was the task of retaining semantics and sensory details used within the lyrics.…”

“…The second problem was the task of retaining semantics and sensory details used within the lyrics. General-purpose translation models might translate the meaning and fluency, but not poetic style when translating poems (5). The same went for the semantics and sensory details when dealing with lyrics.…”

Vers une robotique du traduire ⁠–⁠ Introduction

“…Apparition des autrices et auteurs dans l'ordre alphabétique. 2 Même dans le cas de la poésie, un genre qui présente des défis considérables pour la traduction automatique, des premiers essais de traduction neuronale[Chakrabarty et al, 2021 ; Ghazvininejad et al, 2018] laissent entrevoir le potentiel des systèmes.…”