Poetic Machine Translation by Google
Software engineers at Google have been delving into the world of automatic poetry translation. A paper entitled “Poetic” Statistical Machine Translation: Rhyme and Meter will be presented at the upcoming EMNLP (Empirical Methods in Natural Language Processing) conference and shall address the progress made to date, the difficulties encountered and also discuss the necessary considerations to be taken into account when tackling such a difficult topic.
French and English was the language combination used for the program and there is an option for the user to select the target translation genre (such as sonnet). However, as Google point out, form rather than accuracy is the main focus at present, which unfortunately has a negative ‘impact on translation quality’ as MT is only capable of replicating either form or meaning.
The system is not yet available to the public, but the aforementioned paper gives further details of the work that has already been carried out. The ‘purely technical challenges around generating translations with fixed rhyme and meter schemes’ are discussed and the debate on whether to maintain the form of the source language text in the target translation is also addressed. Translation loss or ‘quality penalty’ when using MT for poetic translation is covered along with stress patterns and poetic form. Sub-sections include line-length, syllables and line breaks; stress and syllables for rhythmic poetry; meter (the ‘exact sequence of stressed and unstressed syllables’) for rhyming constraints; and the importance of avoiding computer-generated ‘self-rhyme’ (identical words used to produce the rhyme).
Even though the paper deals with statistical machine translation and technical issues, it could be said that it has been responsible for producing some original poetry around this subject. The paper was submitted for review and feedback prior to its presentation at the EMNLP conference. However, there was one response that Google had not counted on – one of the reviews was written in verse! This review has been published online, along with further author and reviewer comments – also written in verse!
Google are well aware that the use of MT in this field is certainly in its infancy and the official blog even quotes poet Robert Frost who said ‘Poetry is what gets lost in translation’. Clearly eschewing the use of human translators in this sector will only increase translation loss and misunderstanding, however, it will be interesting to see how these losses can be minimized over time . . . and we’ll be keeping an eye on the Review in Verse, as it really does seem to be a first!
Sources: Google Research Blog (Poetic Machine Translation); “Poetic” Statistical Machine Translation: Rhyme and Meter (Dmitriy Genzel, Jakob Uszkoreit, Franz Och); A Review in Verse: https://research.google.com/archive/papers/review_in_verse.html