The Linguist is a languages magazine for professional linguists, translators, interpreters, language professionals, language teachers, trainers, students and academics with articles on translation, interpreting, business, government, technology
Issue link: https://thelinguist.uberflip.com/i/220935
FOCUS: CAT TOOLS EARLY EXPERIMENT IBM in the 1950s, when the company gave a joint demonstration of machine translation with Georgetown University © ISTOCKPHOTO Post-editing: a new role MT has not just brought interesting research topics and increases in productivity, but also new services. The MT@EU project, which uses the Moses SMT system, was launched in July, and 17 language departments have so far reported that they are satisfied with the quality of the output5 (you've already guessed that Hungarian and Finish were not among them). EU translators are now getting to grips with an important new service: MT post-editing, and especially with the difference between light and full post-editing. (The former is seen as a sacrilegious activity by most professionals, who refuse to believe it is what clients actually want; well, some do!) Like many language service providers (LSPs), the heads of EU language units need to understand more about how much SMT helps. Evaluating the time and effort involved in post-editing is crucial, which is why combining the time spent post-editing with 'edit distance' information (ie, 'the minimum number of single-character edits – insertion, deletion, substitution – required to change one [string] into [an]other'6) is of great interest. At the moment, Kilgray's memoQ 2013 tracks the former, while MemSource, Lingo24's Coach (still in development) and the modified OmegaT – iOmegaT – track both. I welcome such developments. They have already started to show that post-editing does take time and effort (although a MT posteditor is generally faster than a translator); that generic MT output requires more postediting than domain-specific MT output; and that MT works differently for different language combinations and domains. LSPs also need to develop clear postediting instructions for their freelancers (TAUS Vol/52 No/6 2013 has long posted an example7); remunerate this new activity appropriately; and be very careful about passing off MT output for postediting as human translation for revision. Making the leap to MT What could you do in the meantime? The first thing you should know is that MT is only a few clicks away in your preferred CAT tool, ranging from Google and Microsoft Translator (currently free up to 2 million words per month) to a wide range of providers, including Let'sMT! and KantanMT. A popular approach to integrate MT in your workflow would be to pre-translate new texts using your TMs, and then pre-translate the low fuzzy matches and no matches using MT. You can do this easily in pretty much any CAT tool. What you cannot do, though, is see both technologies intertwined, unless your are using Déjà Vu X2, which allows fuzzy match repair either through an online MT system or through its built-in DeepMiner. Should you want to customise (to some extent) an online DIY system without learning to program, Let'sMT! and KantanMT are the best places to start. However, don't assume that any DIY MT system will match the quality of a customised MT system, built using cleaned-up training data (ie, monolingual corpora, translation memories and databases), and constantly refined based on customer feedback. It is akin to learning to swim on your own as opposed to having a coach. If I were to offer any prizes for external MT provider support, memoQ currently would lead the way. However, I would nominate OmegaT for looking after your privacy best, as it only sends the source text to the MT provider – not the accepted or corrected MT output. Alternatively, if you want to build your own system, the free, open-source Moses for Mere Mortals initiative will get you up and running a lot faster than compiling Moses from source code. Whatever option you choose, remember to run realistic tests: this is a tool for technical texts and, ideally, controlled language input, not for transcreating or translating humour. Given the fast-paced nature of the sector, it is important to keep an eye on technological developments and exciting future projects such as CASMACAT and MateCat. These initiatives integrate MT post-editing more seamlessly into the translation workflow, and use machine-learning techniques to improve the quality of MT output. Looking out for new developments will enable you to stay ahead of the game and improve your relationship with clients. If you are an LSP with MT-related research questions, consider forging a strong collaboration with an experienced Translation Studies Centre for mutual benefit. In any case, by keeping an open mind, and using CAT, MT and voice recognition tools, you are getting ready to swim in the fast lane. Notes 1 http://goo.gl/yxTejw 2 http://goo.gl/LgmYjw 3 http://goo.gl/sJCGkU 4 http://goo.gl/MjNXWx 5 Languages and Translation, February 2013, European Commission; http://goo.gl/tjjeHq 6 http://goo.gl/NlWsQu 7 http://goo.gl/ceK6Hv DECEMBER 2013/JANUARY 2014 The Linguist 15