“Poetic” Statistical Machine Translation: Rhyme and Meter

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 158–166, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics “Poetic” Statistical Machine Translation: Rhyme and Meter Dmitriy Genzel Jakob Uszkoreit Franz Och Google, Inc. 1600 Amphitheatre Pkwy Mountain View, CA 94043, USA {dmitriy,uszkoreit,och}@google.com Abstract As a prerequisite to translation of poetry, we implement the ability to produce translations with meter and rhyme for phrase-based MT, examine whether the hypothesis space of such a system is flexible enough to accomodate such constraints, and investigate the impact of such constraints on translation quality. 1 Introduction Machine translation of poetry is probably one of the hardest possible tasks that can be considered in com- putational linguistics, MT, or even AI in general. It is a task that most humans are not truly capable of. Robert Frost is reported to have said that poetry is that which gets lost in translation. Not surprisingly, given the task’s difficulty, we are not aware of any work in the field that attempts to solve this problem, or even discuss it, except to mention its difficulty, and professional translators like to cite it as an exam- ple of an area where MT will never replace a human translator. This may well be true in the near or even long term. However, there are aspects of the prob- lem that we can already tackle, namely the problem of the poetic form. Vladimir Nabokov, in his famous translation of Eugene Onegin (Nabokov, 1965), a poem with a very strict meter and rhyming scheme, heavily dis- parages those translators that attempt to preserve the form, claiming that since it is impossible to perfectly preserve both the form and the meaning, the form must be entirely sacrificed. On the other hand, Dou- glas Hofstadter, who spends 600 pages describing how to translate a 60 word poem in 80 different ways in Le Ton beau de Marot (1998), makes a strong case that a poem’s form must be preserved in translation, if at all possible. Leaving the controversy to the pro- fessional translators, we investigate whether or not it is possible to produce translations that conform to certain metrical constraints common in poetry. Statistical machine translation techniques, unlike their traditional rule-based counterparts, are in fact well-suited to the task. Because the number of po- tential translation hypotheses is very large, it is not unreasonable to expect that some of them should conform to an externally imposed standard. The goal of this paper is to investigate how these hy- potheses can be efficiently identified, how often they are present, and what the quality penalty for impos- ing them is. 2 Related Work There has been very little work related to the transla- tion of poetry. There has been some work where MT techniques were used to produce poetry (Jiang and Zhou, 2008). In other computational poetry work, Ramakrishnan et al (2009) generate song lyrics from melody and various algorithms for poetry gener- ation (Manurung et al., 2000; D´ıaz-Agudo …

Content truncated to 500 words