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

TL;DR

This paper empirically investigates neural poetry translation, exploring data size, style, and multilingual models, and introduces a new poetic translation dataset, showing multilingual fine-tuning on poetic data yields superior results.

Contribution

It provides the first open dataset for poetic translation and systematically compares multilingual and bilingual models, highlighting the effectiveness of poetic fine-tuning.

Findings

Multilingual fine-tuning on poetic data outperforms non-poetic data fine-tuning.

Poetic fine-tuning significantly improves translation quality over larger non-poetic datasets.

Multilingual models outperform bilingual models in 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 zero-shot setup; 2) bilingual vs. multilingual learning; and 3) language-family-specific models vs. mixed-multilingual 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) and human evaluation metrics such as faithfulness (meaning and poetic style). Moreover, multilingual fine-tuning on poetic data outperforms \emph{bilingual} fine-tuning on poetic data.