Towards Zero-Shot Multilingual Poetry Translation

Proceedings of Machine Translation Summit XIX, Vol. 1: Research Track, pages 324–335 September 4–8, 2023, Macau SAR, China. ©2023 The authors. This article is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0). 324 Towards Zero-Shot Multilingual Poetry Translation Wai Lei Song nlp2ct.jacky@gmail.com Haoyun Xu nlp2ct.haoyun@gmail.com Derek F. Wong derekfw@um.edu.mo Runzhe Zhan nlp2ct.runzhe@gmail.com Lidia S. Chao lidiasc@um.edu.mo Shanshan Wang∗ nlp2ct.shanshan@um.edu.mo NLP2CT Lab, Department of Computer and Information Science, University of Macau Abstract The application of machine translation in the field of poetry has always presented significant challenges. Conventional machine translation techniques are inadequate for capturing and translating the unique style of poetry. The absence of a parallel poetry corpus and the dis- tinctive structure of poetry further restrict the effectiveness of traditional methods. This paper introduces a zero-shot method that is capable of translating poetry style without the need for a large-scale training corpus. Specifically, we treat poetry translation as a standard machine translation problem and subsequently inject the poetry style upon completion of the translation process. Our injection model only requires back-translation and easily obtainable monolingual data, making it a low-cost solution. We conducted experiments on three translation directions and presented automatic and human evaluations, demonstrating that our proposed method out- performs existing online systems and other competitive baselines. These results validate the feasibility and potential of our proposed approach and provide new prospects for poetry trans- lation. 1 Introduction The process of translating poetry presents an intricate challenge within the field of machine translation. The prevalence of the neural machine translation (NMT) paradigm (Vaswani et al., 2017), which necessitates copious amounts of data to effectively train a model capable of pro- ducing accurate translations (Koehn and Knowles, 2017). Unfortunately, the availability of parallel corpora that can be leveraged towards the training of a robust poetry translation system is currently inadequate. On the other hand, poetry is a manifestation of the unique imagination and creativity of the poet, as well as their distinctive writing style. As Chakrabarty et al. (2021) has pointed out, although NMT systems may succeed in translating the essence meaning of the poetry, the translation process inevitably disregards the writing style. As a result, two distinct research paths have emerged within the field of poetry transla- tion. In an effort to preserve the poet’s distinctive writing style, Genzel et al. (2010) initially employed statistical machine translation, with a focus on maintaining the rhythm of the original ∗Corresponding Author. 325 Poetry (Source Language) Ordinary Translate Style Injection Back-translation Iterative Back-translation General Text (Target Language) Poetry (Target Language) General Text (Target Language) Poetry (Target Language) Trained by Figure 1: An illustration of the proposed zero-shot poetry translation method. work. This approach was subsequently refined by Ghazvininejad et al. (2018), who introduced the additional constraints of rhythm and rhyme into neural poetry translation. This yielded an acceptability rate of 78.2% in terms of translation quality. Additionally, Yang et al. (2019) ex- plored the use of unique tokens to govern the structure of the translated poetry. Meanwhile, …

Content truncated to 500 words