Crawl and crowd to bring machine translation to under-resourced languages

Toral, Antonio; Esplà-Gomis, Miquel; Klubička, Filip; Ljubešić, Nikola; Papavassiliou, Vassilis; Prokopidis, Prokopis; Rubino, Raphaël; Way, Andy

doi:10.1007/s10579-016-9363-6

Cited by 7 publications

(2 citation statements)

References 22 publications

(21 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In common with other applications of machine learning, human-created data is required for training in vast quantities. Webcrawling for parallel text is a cost-effective option for datagathering and may be useful when high-quality data is scarce (Toral et al 2017). However, neural networks ideally require large amounts of high-quality training data and this is particularly true of NMT since the popularization of the system architecture known as the transformer model, which has exhibited efficiency gains (Vaswani et al 2017).…”

Section: Copyright and Machine Learningmentioning

confidence: 99%

“…i This is, of course, untrue. Notwithstanding the possibility of webcrawling for parallel texts as a method of gathering MT training data (Toral et al 2017), the most valuable and highest-quality data for training MT systems, as with other applications of machine learning based on human data, is aligned human translations, in which source and target texts are computationally linked, usually at the sentence level.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Copyright and the re-use of translation as data

Moorkens¹,

Lewis²

2019

The Routledge Handbook of Translation and Technology

View full text Add to dashboard Cite

Translation copyright was first codified as a derivative work in the Berne Convention of 1886, subject to the rights of the creator of the original work. While most countries are now contracting parties to the Berne Convention, differing interpretations and additional laws and directives mean that rights to ownership of a translation are not consistent in all jurisdictions. The original intention of the Berne Convention was to protect authors' rights and to prevent piracy, and the authors could not have foreseen the large scale reuse of translations, initially via translation memory tools, then as training data for machine translation (MT) systems. Parallel data is repurposed in ever-increasing amounts, but broken down to word and subword levels. At present, rights to ownership are rarely passed to the translator, meaning that, while an initial translation may be costly, secondary uses are very inexpensive. This chapter explores the history of translation copyright and leveraging, and introduces concerns relating to machine learning more generally and applies them to translation.

show abstract

Section: Copyright and Machine Learningmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%